Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * bufmgr.c
4 : * buffer manager interface routines
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/storage/buffer/bufmgr.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : /*
16 : * Principal entry points:
17 : *
18 : * ReadBuffer() -- find or create a buffer holding the requested page,
19 : * and pin it so that no one can destroy it while this process
20 : * is using it.
21 : *
22 : * StartReadBuffer() -- as above, with separate wait step
23 : * StartReadBuffers() -- multiple block version
24 : * WaitReadBuffers() -- second step of above
25 : *
26 : * ReleaseBuffer() -- unpin a buffer
27 : *
28 : * MarkBufferDirty() -- mark a pinned buffer's contents as "dirty".
29 : * The disk write is delayed until buffer replacement or checkpoint.
30 : *
31 : * See also these files:
32 : * freelist.c -- chooses victim for buffer replacement
33 : * buf_table.c -- manages the buffer lookup table
34 : */
35 : #include "postgres.h"
36 :
37 : #include <sys/file.h>
38 : #include <unistd.h>
39 :
40 : #include "access/tableam.h"
41 : #include "access/xloginsert.h"
42 : #include "access/xlogutils.h"
43 : #ifdef USE_ASSERT_CHECKING
44 : #include "catalog/pg_tablespace_d.h"
45 : #endif
46 : #include "catalog/storage.h"
47 : #include "catalog/storage_xlog.h"
48 : #include "executor/instrument.h"
49 : #include "lib/binaryheap.h"
50 : #include "miscadmin.h"
51 : #include "pg_trace.h"
52 : #include "pgstat.h"
53 : #include "postmaster/bgwriter.h"
54 : #include "storage/aio.h"
55 : #include "storage/buf_internals.h"
56 : #include "storage/bufmgr.h"
57 : #include "storage/fd.h"
58 : #include "storage/ipc.h"
59 : #include "storage/lmgr.h"
60 : #include "storage/proc.h"
61 : #include "storage/read_stream.h"
62 : #include "storage/smgr.h"
63 : #include "storage/standby.h"
64 : #include "utils/memdebug.h"
65 : #include "utils/ps_status.h"
66 : #include "utils/rel.h"
67 : #include "utils/resowner.h"
68 : #include "utils/timestamp.h"
69 :
70 :
71 : /* Note: these two macros only work on shared buffers, not local ones! */
72 : #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
73 : #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
74 :
75 : /* Note: this macro only works on local buffers, not shared ones! */
76 : #define LocalBufHdrGetBlock(bufHdr) \
77 : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
78 :
79 : /* Bits in SyncOneBuffer's return value */
80 : #define BUF_WRITTEN 0x01
81 : #define BUF_REUSABLE 0x02
82 :
83 : #define RELS_BSEARCH_THRESHOLD 20
84 :
85 : /*
86 : * This is the size (in the number of blocks) above which we scan the
87 : * entire buffer pool to remove the buffers for all the pages of relation
88 : * being dropped. For the relations with size below this threshold, we find
89 : * the buffers by doing lookups in BufMapping table.
90 : */
91 : #define BUF_DROP_FULL_SCAN_THRESHOLD (uint64) (NBuffers / 32)
92 :
93 : typedef struct PrivateRefCountEntry
94 : {
95 : Buffer buffer;
96 : int32 refcount;
97 : } PrivateRefCountEntry;
98 :
99 : /* 64 bytes, about the size of a cache line on common systems */
100 : #define REFCOUNT_ARRAY_ENTRIES 8
101 :
102 : /*
103 : * Status of buffers to checkpoint for a particular tablespace, used
104 : * internally in BufferSync.
105 : */
106 : typedef struct CkptTsStatus
107 : {
108 : /* oid of the tablespace */
109 : Oid tsId;
110 :
111 : /*
112 : * Checkpoint progress for this tablespace. To make progress comparable
113 : * between tablespaces the progress is, for each tablespace, measured as a
114 : * number between 0 and the total number of to-be-checkpointed pages. Each
115 : * page checkpointed in this tablespace increments this space's progress
116 : * by progress_slice.
117 : */
118 : float8 progress;
119 : float8 progress_slice;
120 :
121 : /* number of to-be checkpointed pages in this tablespace */
122 : int num_to_scan;
123 : /* already processed pages in this tablespace */
124 : int num_scanned;
125 :
126 : /* current offset in CkptBufferIds for this tablespace */
127 : int index;
128 : } CkptTsStatus;
129 :
130 : /*
131 : * Type for array used to sort SMgrRelations
132 : *
133 : * FlushRelationsAllBuffers shares the same comparator function with
134 : * DropRelationsAllBuffers. Pointer to this struct and RelFileLocator must be
135 : * compatible.
136 : */
137 : typedef struct SMgrSortArray
138 : {
139 : RelFileLocator rlocator; /* This must be the first member */
140 : SMgrRelation srel;
141 : } SMgrSortArray;
142 :
143 : /* GUC variables */
144 : bool zero_damaged_pages = false;
145 : int bgwriter_lru_maxpages = 100;
146 : double bgwriter_lru_multiplier = 2.0;
147 : bool track_io_timing = false;
148 :
149 : /*
150 : * How many buffers PrefetchBuffer callers should try to stay ahead of their
151 : * ReadBuffer calls by. Zero means "never prefetch". This value is only used
152 : * for buffers not belonging to tablespaces that have their
153 : * effective_io_concurrency parameter set.
154 : */
155 : int effective_io_concurrency = DEFAULT_EFFECTIVE_IO_CONCURRENCY;
156 :
157 : /*
158 : * Like effective_io_concurrency, but used by maintenance code paths that might
159 : * benefit from a higher setting because they work on behalf of many sessions.
160 : * Overridden by the tablespace setting of the same name.
161 : */
162 : int maintenance_io_concurrency = DEFAULT_MAINTENANCE_IO_CONCURRENCY;
163 :
164 : /*
165 : * Limit on how many blocks should be handled in single I/O operations.
166 : * StartReadBuffers() callers should respect it, as should other operations
167 : * that call smgr APIs directly. It is computed as the minimum of underlying
168 : * GUCs io_combine_limit_guc and io_max_combine_limit.
169 : */
170 : int io_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
171 : int io_combine_limit_guc = DEFAULT_IO_COMBINE_LIMIT;
172 : int io_max_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
173 :
174 : /*
175 : * GUC variables about triggering kernel writeback for buffers written; OS
176 : * dependent defaults are set via the GUC mechanism.
177 : */
178 : int checkpoint_flush_after = DEFAULT_CHECKPOINT_FLUSH_AFTER;
179 : int bgwriter_flush_after = DEFAULT_BGWRITER_FLUSH_AFTER;
180 : int backend_flush_after = DEFAULT_BACKEND_FLUSH_AFTER;
181 :
182 : /* local state for LockBufferForCleanup */
183 : static BufferDesc *PinCountWaitBuf = NULL;
184 :
185 : /*
186 : * Backend-Private refcount management:
187 : *
188 : * Each buffer also has a private refcount that keeps track of the number of
189 : * times the buffer is pinned in the current process. This is so that the
190 : * shared refcount needs to be modified only once if a buffer is pinned more
191 : * than once by an individual backend. It's also used to check that no buffers
192 : * are still pinned at the end of transactions and when exiting.
193 : *
194 : *
195 : * To avoid - as we used to - requiring an array with NBuffers entries to keep
196 : * track of local buffers, we use a small sequentially searched array
197 : * (PrivateRefCountArray) and an overflow hash table (PrivateRefCountHash) to
198 : * keep track of backend local pins.
199 : *
200 : * Until no more than REFCOUNT_ARRAY_ENTRIES buffers are pinned at once, all
201 : * refcounts are kept track of in the array; after that, new array entries
202 : * displace old ones into the hash table. That way a frequently used entry
203 : * can't get "stuck" in the hashtable while infrequent ones clog the array.
204 : *
205 : * Note that in most scenarios the number of pinned buffers will not exceed
206 : * REFCOUNT_ARRAY_ENTRIES.
207 : *
208 : *
209 : * To enter a buffer into the refcount tracking mechanism first reserve a free
210 : * entry using ReservePrivateRefCountEntry() and then later, if necessary,
211 : * fill it with NewPrivateRefCountEntry(). That split lets us avoid doing
212 : * memory allocations in NewPrivateRefCountEntry() which can be important
213 : * because in some scenarios it's called with a spinlock held...
214 : */
215 : static struct PrivateRefCountEntry PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES];
216 : static HTAB *PrivateRefCountHash = NULL;
217 : static int32 PrivateRefCountOverflowed = 0;
218 : static uint32 PrivateRefCountClock = 0;
219 : static PrivateRefCountEntry *ReservedRefCountEntry = NULL;
220 :
221 : static uint32 MaxProportionalPins;
222 :
223 : static void ReservePrivateRefCountEntry(void);
224 : static PrivateRefCountEntry *NewPrivateRefCountEntry(Buffer buffer);
225 : static PrivateRefCountEntry *GetPrivateRefCountEntry(Buffer buffer, bool do_move);
226 : static inline int32 GetPrivateRefCount(Buffer buffer);
227 : static void ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref);
228 :
229 : /* ResourceOwner callbacks to hold in-progress I/Os and buffer pins */
230 : static void ResOwnerReleaseBufferIO(Datum res);
231 : static char *ResOwnerPrintBufferIO(Datum res);
232 : static void ResOwnerReleaseBufferPin(Datum res);
233 : static char *ResOwnerPrintBufferPin(Datum res);
234 :
235 : const ResourceOwnerDesc buffer_io_resowner_desc =
236 : {
237 : .name = "buffer io",
238 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
239 : .release_priority = RELEASE_PRIO_BUFFER_IOS,
240 : .ReleaseResource = ResOwnerReleaseBufferIO,
241 : .DebugPrint = ResOwnerPrintBufferIO
242 : };
243 :
244 : const ResourceOwnerDesc buffer_pin_resowner_desc =
245 : {
246 : .name = "buffer pin",
247 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
248 : .release_priority = RELEASE_PRIO_BUFFER_PINS,
249 : .ReleaseResource = ResOwnerReleaseBufferPin,
250 : .DebugPrint = ResOwnerPrintBufferPin
251 : };
252 :
253 : /*
254 : * Ensure that the PrivateRefCountArray has sufficient space to store one more
255 : * entry. This has to be called before using NewPrivateRefCountEntry() to fill
256 : * a new entry - but it's perfectly fine to not use a reserved entry.
257 : */
258 : static void
259 130231396 : ReservePrivateRefCountEntry(void)
260 : {
261 : /* Already reserved (or freed), nothing to do */
262 130231396 : if (ReservedRefCountEntry != NULL)
263 122027248 : return;
264 :
265 : /*
266 : * First search for a free entry the array, that'll be sufficient in the
267 : * majority of cases.
268 : */
269 : {
270 : int i;
271 :
272 20925706 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
273 : {
274 : PrivateRefCountEntry *res;
275 :
276 20582844 : res = &PrivateRefCountArray[i];
277 :
278 20582844 : if (res->buffer == InvalidBuffer)
279 : {
280 7861286 : ReservedRefCountEntry = res;
281 7861286 : return;
282 : }
283 : }
284 : }
285 :
286 : /*
287 : * No luck. All array entries are full. Move one array entry into the hash
288 : * table.
289 : */
290 : {
291 : /*
292 : * Move entry from the current clock position in the array into the
293 : * hashtable. Use that slot.
294 : */
295 : PrivateRefCountEntry *hashent;
296 : bool found;
297 :
298 : /* select victim slot */
299 342862 : ReservedRefCountEntry =
300 342862 : &PrivateRefCountArray[PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES];
301 :
302 : /* Better be used, otherwise we shouldn't get here. */
303 : Assert(ReservedRefCountEntry->buffer != InvalidBuffer);
304 :
305 : /* enter victim array entry into hashtable */
306 342862 : hashent = hash_search(PrivateRefCountHash,
307 342862 : &(ReservedRefCountEntry->buffer),
308 : HASH_ENTER,
309 : &found);
310 : Assert(!found);
311 342862 : hashent->refcount = ReservedRefCountEntry->refcount;
312 :
313 : /* clear the now free array slot */
314 342862 : ReservedRefCountEntry->buffer = InvalidBuffer;
315 342862 : ReservedRefCountEntry->refcount = 0;
316 :
317 342862 : PrivateRefCountOverflowed++;
318 : }
319 : }
320 :
321 : /*
322 : * Fill a previously reserved refcount entry.
323 : */
324 : static PrivateRefCountEntry *
325 118581762 : NewPrivateRefCountEntry(Buffer buffer)
326 : {
327 : PrivateRefCountEntry *res;
328 :
329 : /* only allowed to be called when a reservation has been made */
330 : Assert(ReservedRefCountEntry != NULL);
331 :
332 : /* use up the reserved entry */
333 118581762 : res = ReservedRefCountEntry;
334 118581762 : ReservedRefCountEntry = NULL;
335 :
336 : /* and fill it */
337 118581762 : res->buffer = buffer;
338 118581762 : res->refcount = 0;
339 :
340 118581762 : return res;
341 : }
342 :
343 : /*
344 : * Return the PrivateRefCount entry for the passed buffer.
345 : *
346 : * Returns NULL if a buffer doesn't have a refcount entry. Otherwise, if
347 : * do_move is true, and the entry resides in the hashtable the entry is
348 : * optimized for frequent access by moving it to the array.
349 : */
350 : static PrivateRefCountEntry *
351 293118398 : GetPrivateRefCountEntry(Buffer buffer, bool do_move)
352 : {
353 : PrivateRefCountEntry *res;
354 : int i;
355 :
356 : Assert(BufferIsValid(buffer));
357 : Assert(!BufferIsLocal(buffer));
358 :
359 : /*
360 : * First search for references in the array, that'll be sufficient in the
361 : * majority of cases.
362 : */
363 1380692852 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
364 : {
365 1266095954 : res = &PrivateRefCountArray[i];
366 :
367 1266095954 : if (res->buffer == buffer)
368 178521500 : return res;
369 : }
370 :
371 : /*
372 : * By here we know that the buffer, if already pinned, isn't residing in
373 : * the array.
374 : *
375 : * Only look up the buffer in the hashtable if we've previously overflowed
376 : * into it.
377 : */
378 114596898 : if (PrivateRefCountOverflowed == 0)
379 113768088 : return NULL;
380 :
381 828810 : res = hash_search(PrivateRefCountHash, &buffer, HASH_FIND, NULL);
382 :
383 828810 : if (res == NULL)
384 481142 : return NULL;
385 347668 : else if (!do_move)
386 : {
387 : /* caller doesn't want us to move the hash entry into the array */
388 305130 : return res;
389 : }
390 : else
391 : {
392 : /* move buffer from hashtable into the free array slot */
393 : bool found;
394 : PrivateRefCountEntry *free;
395 :
396 : /* Ensure there's a free array slot */
397 42538 : ReservePrivateRefCountEntry();
398 :
399 : /* Use up the reserved slot */
400 : Assert(ReservedRefCountEntry != NULL);
401 42538 : free = ReservedRefCountEntry;
402 42538 : ReservedRefCountEntry = NULL;
403 : Assert(free->buffer == InvalidBuffer);
404 :
405 : /* and fill it */
406 42538 : free->buffer = buffer;
407 42538 : free->refcount = res->refcount;
408 :
409 : /* delete from hashtable */
410 42538 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
411 : Assert(found);
412 : Assert(PrivateRefCountOverflowed > 0);
413 42538 : PrivateRefCountOverflowed--;
414 :
415 42538 : return free;
416 : }
417 : }
418 :
419 : /*
420 : * Returns how many times the passed buffer is pinned by this backend.
421 : *
422 : * Only works for shared memory buffers!
423 : */
424 : static inline int32
425 6427582 : GetPrivateRefCount(Buffer buffer)
426 : {
427 : PrivateRefCountEntry *ref;
428 :
429 : Assert(BufferIsValid(buffer));
430 : Assert(!BufferIsLocal(buffer));
431 :
432 : /*
433 : * Not moving the entry - that's ok for the current users, but we might
434 : * want to change this one day.
435 : */
436 6427582 : ref = GetPrivateRefCountEntry(buffer, false);
437 :
438 6427582 : if (ref == NULL)
439 124096 : return 0;
440 6303486 : return ref->refcount;
441 : }
442 :
443 : /*
444 : * Release resources used to track the reference count of a buffer which we no
445 : * longer have pinned and don't want to pin again immediately.
446 : */
447 : static void
448 118581762 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
449 : {
450 : Assert(ref->refcount == 0);
451 :
452 118581762 : if (ref >= &PrivateRefCountArray[0] &&
453 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
454 : {
455 118281438 : ref->buffer = InvalidBuffer;
456 :
457 : /*
458 : * Mark the just used entry as reserved - in many scenarios that
459 : * allows us to avoid ever having to search the array/hash for free
460 : * entries.
461 : */
462 118281438 : ReservedRefCountEntry = ref;
463 : }
464 : else
465 : {
466 : bool found;
467 300324 : Buffer buffer = ref->buffer;
468 :
469 300324 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
470 : Assert(found);
471 : Assert(PrivateRefCountOverflowed > 0);
472 300324 : PrivateRefCountOverflowed--;
473 : }
474 118581762 : }
475 :
476 : /*
477 : * BufferIsPinned
478 : * True iff the buffer is pinned (also checks for valid buffer number).
479 : *
480 : * NOTE: what we check here is that *this* backend holds a pin on
481 : * the buffer. We do not care whether some other backend does.
482 : */
483 : #define BufferIsPinned(bufnum) \
484 : ( \
485 : !BufferIsValid(bufnum) ? \
486 : false \
487 : : \
488 : BufferIsLocal(bufnum) ? \
489 : (LocalRefCount[-(bufnum) - 1] > 0) \
490 : : \
491 : (GetPrivateRefCount(bufnum) > 0) \
492 : )
493 :
494 :
495 : static Buffer ReadBuffer_common(Relation rel,
496 : SMgrRelation smgr, char smgr_persistence,
497 : ForkNumber forkNum, BlockNumber blockNum,
498 : ReadBufferMode mode, BufferAccessStrategy strategy);
499 : static BlockNumber ExtendBufferedRelCommon(BufferManagerRelation bmr,
500 : ForkNumber fork,
501 : BufferAccessStrategy strategy,
502 : uint32 flags,
503 : uint32 extend_by,
504 : BlockNumber extend_upto,
505 : Buffer *buffers,
506 : uint32 *extended_by);
507 : static BlockNumber ExtendBufferedRelShared(BufferManagerRelation bmr,
508 : ForkNumber fork,
509 : BufferAccessStrategy strategy,
510 : uint32 flags,
511 : uint32 extend_by,
512 : BlockNumber extend_upto,
513 : Buffer *buffers,
514 : uint32 *extended_by);
515 : static bool PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
516 : bool skip_if_not_valid);
517 : static void PinBuffer_Locked(BufferDesc *buf);
518 : static void UnpinBuffer(BufferDesc *buf);
519 : static void UnpinBufferNoOwner(BufferDesc *buf);
520 : static void BufferSync(int flags);
521 : static int SyncOneBuffer(int buf_id, bool skip_recently_used,
522 : WritebackContext *wb_context);
523 : static void WaitIO(BufferDesc *buf);
524 : static void AbortBufferIO(Buffer buffer);
525 : static void shared_buffer_write_error_callback(void *arg);
526 : static void local_buffer_write_error_callback(void *arg);
527 : static inline BufferDesc *BufferAlloc(SMgrRelation smgr,
528 : char relpersistence,
529 : ForkNumber forkNum,
530 : BlockNumber blockNum,
531 : BufferAccessStrategy strategy,
532 : bool *foundPtr, IOContext io_context);
533 : static bool AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress);
534 : static void CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete);
535 : static Buffer GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context);
536 : static void FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
537 : IOObject io_object, IOContext io_context);
538 : static void FlushBuffer(BufferDesc *buf, SMgrRelation reln,
539 : IOObject io_object, IOContext io_context);
540 : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
541 : ForkNumber forkNum,
542 : BlockNumber nForkBlock,
543 : BlockNumber firstDelBlock);
544 : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
545 : RelFileLocator dstlocator,
546 : ForkNumber forkNum, bool permanent);
547 : static void AtProcExit_Buffers(int code, Datum arg);
548 : static void CheckForBufferLeaks(void);
549 : #ifdef USE_ASSERT_CHECKING
550 : static void AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
551 : void *unused_context);
552 : #endif
553 : static int rlocator_comparator(const void *p1, const void *p2);
554 : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
555 : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
556 : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
557 :
558 :
559 : /*
560 : * Implementation of PrefetchBuffer() for shared buffers.
561 : */
562 : PrefetchBufferResult
563 62202 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
564 : ForkNumber forkNum,
565 : BlockNumber blockNum)
566 : {
567 62202 : PrefetchBufferResult result = {InvalidBuffer, false};
568 : BufferTag newTag; /* identity of requested block */
569 : uint32 newHash; /* hash value for newTag */
570 : LWLock *newPartitionLock; /* buffer partition lock for it */
571 : int buf_id;
572 :
573 : Assert(BlockNumberIsValid(blockNum));
574 :
575 : /* create a tag so we can lookup the buffer */
576 62202 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
577 : forkNum, blockNum);
578 :
579 : /* determine its hash code and partition lock ID */
580 62202 : newHash = BufTableHashCode(&newTag);
581 62202 : newPartitionLock = BufMappingPartitionLock(newHash);
582 :
583 : /* see if the block is in the buffer pool already */
584 62202 : LWLockAcquire(newPartitionLock, LW_SHARED);
585 62202 : buf_id = BufTableLookup(&newTag, newHash);
586 62202 : LWLockRelease(newPartitionLock);
587 :
588 : /* If not in buffers, initiate prefetch */
589 62202 : if (buf_id < 0)
590 : {
591 : #ifdef USE_PREFETCH
592 : /*
593 : * Try to initiate an asynchronous read. This returns false in
594 : * recovery if the relation file doesn't exist.
595 : */
596 34716 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
597 17134 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
598 : {
599 17134 : result.initiated_io = true;
600 : }
601 : #endif /* USE_PREFETCH */
602 : }
603 : else
604 : {
605 : /*
606 : * Report the buffer it was in at that time. The caller may be able
607 : * to avoid a buffer table lookup, but it's not pinned and it must be
608 : * rechecked!
609 : */
610 44620 : result.recent_buffer = buf_id + 1;
611 : }
612 :
613 : /*
614 : * If the block *is* in buffers, we do nothing. This is not really ideal:
615 : * the block might be just about to be evicted, which would be stupid
616 : * since we know we are going to need it soon. But the only easy answer
617 : * is to bump the usage_count, which does not seem like a great solution:
618 : * when the caller does ultimately touch the block, usage_count would get
619 : * bumped again, resulting in too much favoritism for blocks that are
620 : * involved in a prefetch sequence. A real fix would involve some
621 : * additional per-buffer state, and it's not clear that there's enough of
622 : * a problem to justify that.
623 : */
624 :
625 62202 : return result;
626 : }
627 :
628 : /*
629 : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
630 : *
631 : * This is named by analogy to ReadBuffer but doesn't actually allocate a
632 : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
633 : * block will not be delayed by the I/O. Prefetching is optional.
634 : *
635 : * There are three possible outcomes:
636 : *
637 : * 1. If the block is already cached, the result includes a valid buffer that
638 : * could be used by the caller to avoid the need for a later buffer lookup, but
639 : * it's not pinned, so the caller must recheck it.
640 : *
641 : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
642 : * true. Currently there is no way to know if the data was already cached by
643 : * the kernel and therefore didn't really initiate I/O, and no way to know when
644 : * the I/O completes other than using synchronous ReadBuffer().
645 : *
646 : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
647 : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
648 : * lack of a kernel facility), direct I/O is enabled, or the underlying
649 : * relation file wasn't found and we are in recovery. (If the relation file
650 : * wasn't found and we are not in recovery, an error is raised).
651 : */
652 : PrefetchBufferResult
653 40678 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
654 : {
655 : Assert(RelationIsValid(reln));
656 : Assert(BlockNumberIsValid(blockNum));
657 :
658 40678 : if (RelationUsesLocalBuffers(reln))
659 : {
660 : /* see comments in ReadBufferExtended */
661 1566 : if (RELATION_IS_OTHER_TEMP(reln))
662 0 : ereport(ERROR,
663 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
664 : errmsg("cannot access temporary tables of other sessions")));
665 :
666 : /* pass it off to localbuf.c */
667 1566 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
668 : }
669 : else
670 : {
671 : /* pass it to the shared buffer version */
672 39112 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
673 : }
674 : }
675 :
676 : /*
677 : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
678 : *
679 : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
680 : * successful. Return true if the buffer is valid and still has the expected
681 : * tag. In that case, the buffer is pinned and the usage count is bumped.
682 : */
683 : bool
684 9366 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
685 : Buffer recent_buffer)
686 : {
687 : BufferDesc *bufHdr;
688 : BufferTag tag;
689 : uint32 buf_state;
690 :
691 : Assert(BufferIsValid(recent_buffer));
692 :
693 9366 : ResourceOwnerEnlarge(CurrentResourceOwner);
694 9366 : ReservePrivateRefCountEntry();
695 9366 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
696 :
697 9366 : if (BufferIsLocal(recent_buffer))
698 : {
699 64 : int b = -recent_buffer - 1;
700 :
701 64 : bufHdr = GetLocalBufferDescriptor(b);
702 64 : buf_state = pg_atomic_read_u32(&bufHdr->state);
703 :
704 : /* Is it still valid and holding the right tag? */
705 64 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
706 : {
707 64 : PinLocalBuffer(bufHdr, true);
708 :
709 64 : pgBufferUsage.local_blks_hit++;
710 :
711 64 : return true;
712 : }
713 : }
714 : else
715 : {
716 9302 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
717 :
718 : /*
719 : * Is it still valid and holding the right tag? We do an unlocked tag
720 : * comparison first, to make it unlikely that we'll increment the
721 : * usage counter of the wrong buffer, if someone calls us with a very
722 : * out of date recent_buffer. Then we'll check it again if we get the
723 : * pin.
724 : */
725 18538 : if (BufferTagsEqual(&tag, &bufHdr->tag) &&
726 9236 : PinBuffer(bufHdr, NULL, true))
727 : {
728 9224 : if (BufferTagsEqual(&tag, &bufHdr->tag))
729 : {
730 9224 : pgBufferUsage.shared_blks_hit++;
731 9224 : return true;
732 : }
733 0 : UnpinBuffer(bufHdr);
734 : }
735 : }
736 :
737 78 : return false;
738 : }
739 :
740 : /*
741 : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
742 : * fork with RBM_NORMAL mode and default strategy.
743 : */
744 : Buffer
745 86725418 : ReadBuffer(Relation reln, BlockNumber blockNum)
746 : {
747 86725418 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
748 : }
749 :
750 : /*
751 : * ReadBufferExtended -- returns a buffer containing the requested
752 : * block of the requested relation. If the blknum
753 : * requested is P_NEW, extend the relation file and
754 : * allocate a new block. (Caller is responsible for
755 : * ensuring that only one backend tries to extend a
756 : * relation at the same time!)
757 : *
758 : * Returns: the buffer number for the buffer containing
759 : * the block read. The returned buffer has been pinned.
760 : * Does not return on error --- elog's instead.
761 : *
762 : * Assume when this function is called, that reln has been opened already.
763 : *
764 : * In RBM_NORMAL mode, the page is read from disk, and the page header is
765 : * validated. An error is thrown if the page header is not valid. (But
766 : * note that an all-zero page is considered "valid"; see
767 : * PageIsVerified().)
768 : *
769 : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
770 : * valid, the page is zeroed instead of throwing an error. This is intended
771 : * for non-critical data, where the caller is prepared to repair errors.
772 : *
773 : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
774 : * filled with zeros instead of reading it from disk. Useful when the caller
775 : * is going to fill the page from scratch, since this saves I/O and avoids
776 : * unnecessary failure if the page-on-disk has corrupt page headers.
777 : * The page is returned locked to ensure that the caller has a chance to
778 : * initialize the page before it's made visible to others.
779 : * Caution: do not use this mode to read a page that is beyond the relation's
780 : * current physical EOF; that is likely to cause problems in md.c when
781 : * the page is modified and written out. P_NEW is OK, though.
782 : *
783 : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
784 : * a cleanup-strength lock on the page.
785 : *
786 : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
787 : *
788 : * If strategy is not NULL, a nondefault buffer access strategy is used.
789 : * See buffer/README for details.
790 : */
791 : inline Buffer
792 104732302 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
793 : ReadBufferMode mode, BufferAccessStrategy strategy)
794 : {
795 : Buffer buf;
796 :
797 : /*
798 : * Reject attempts to read non-local temporary relations; we would be
799 : * likely to get wrong data since we have no visibility into the owning
800 : * session's local buffers.
801 : */
802 104732302 : if (RELATION_IS_OTHER_TEMP(reln))
803 0 : ereport(ERROR,
804 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
805 : errmsg("cannot access temporary tables of other sessions")));
806 :
807 : /*
808 : * Read the buffer, and update pgstat counters to reflect a cache hit or
809 : * miss.
810 : */
811 104732302 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
812 : forkNum, blockNum, mode, strategy);
813 :
814 104732256 : return buf;
815 : }
816 :
817 :
818 : /*
819 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
820 : * a relcache entry for the relation.
821 : *
822 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
823 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
824 : * cannot be used for temporary relations (and making that work might be
825 : * difficult, unless we only want to read temporary relations for our own
826 : * ProcNumber).
827 : */
828 : Buffer
829 11373098 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
830 : BlockNumber blockNum, ReadBufferMode mode,
831 : BufferAccessStrategy strategy, bool permanent)
832 : {
833 11373098 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
834 :
835 11373098 : return ReadBuffer_common(NULL, smgr,
836 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
837 : forkNum, blockNum,
838 : mode, strategy);
839 : }
840 :
841 : /*
842 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
843 : */
844 : Buffer
845 90980 : ExtendBufferedRel(BufferManagerRelation bmr,
846 : ForkNumber forkNum,
847 : BufferAccessStrategy strategy,
848 : uint32 flags)
849 : {
850 : Buffer buf;
851 90980 : uint32 extend_by = 1;
852 :
853 90980 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
854 : &buf, &extend_by);
855 :
856 90980 : return buf;
857 : }
858 :
859 : /*
860 : * Extend relation by multiple blocks.
861 : *
862 : * Tries to extend the relation by extend_by blocks. Depending on the
863 : * availability of resources the relation may end up being extended by a
864 : * smaller number of pages (unless an error is thrown, always by at least one
865 : * page). *extended_by is updated to the number of pages the relation has been
866 : * extended to.
867 : *
868 : * buffers needs to be an array that is at least extend_by long. Upon
869 : * completion, the first extend_by array elements will point to a pinned
870 : * buffer.
871 : *
872 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
873 : * locked. This is useful for callers that want a buffer that is guaranteed to
874 : * be empty.
875 : */
876 : BlockNumber
877 319850 : ExtendBufferedRelBy(BufferManagerRelation bmr,
878 : ForkNumber fork,
879 : BufferAccessStrategy strategy,
880 : uint32 flags,
881 : uint32 extend_by,
882 : Buffer *buffers,
883 : uint32 *extended_by)
884 : {
885 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
886 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
887 : Assert(extend_by > 0);
888 :
889 319850 : if (bmr.relpersistence == '\0')
890 319850 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
891 :
892 319850 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
893 : extend_by, InvalidBlockNumber,
894 : buffers, extended_by);
895 : }
896 :
897 : /*
898 : * Extend the relation so it is at least extend_to blocks large, return buffer
899 : * (extend_to - 1).
900 : *
901 : * This is useful for callers that want to write a specific page, regardless
902 : * of the current size of the relation (e.g. useful for visibilitymap and for
903 : * crash recovery).
904 : */
905 : Buffer
906 102996 : ExtendBufferedRelTo(BufferManagerRelation bmr,
907 : ForkNumber fork,
908 : BufferAccessStrategy strategy,
909 : uint32 flags,
910 : BlockNumber extend_to,
911 : ReadBufferMode mode)
912 : {
913 : BlockNumber current_size;
914 102996 : uint32 extended_by = 0;
915 102996 : Buffer buffer = InvalidBuffer;
916 : Buffer buffers[64];
917 :
918 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
919 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
920 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
921 :
922 102996 : if (bmr.relpersistence == '\0')
923 14104 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
924 :
925 : /*
926 : * If desired, create the file if it doesn't exist. If
927 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
928 : * an smgrexists call.
929 : */
930 102996 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
931 14104 : (BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == 0 ||
932 42 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
933 14062 : !smgrexists(BMR_GET_SMGR(bmr), fork))
934 : {
935 14040 : LockRelationForExtension(bmr.rel, ExclusiveLock);
936 :
937 : /* recheck, fork might have been created concurrently */
938 14040 : if (!smgrexists(BMR_GET_SMGR(bmr), fork))
939 14030 : smgrcreate(BMR_GET_SMGR(bmr), fork, flags & EB_PERFORMING_RECOVERY);
940 :
941 14040 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
942 : }
943 :
944 : /*
945 : * If requested, invalidate size cache, so that smgrnblocks asks the
946 : * kernel.
947 : */
948 102996 : if (flags & EB_CLEAR_SIZE_CACHE)
949 14104 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
950 :
951 : /*
952 : * Estimate how many pages we'll need to extend by. This avoids acquiring
953 : * unnecessarily many victim buffers.
954 : */
955 102996 : current_size = smgrnblocks(BMR_GET_SMGR(bmr), fork);
956 :
957 : /*
958 : * Since no-one else can be looking at the page contents yet, there is no
959 : * difference between an exclusive lock and a cleanup-strength lock. Note
960 : * that we pass the original mode to ReadBuffer_common() below, when
961 : * falling back to reading the buffer to a concurrent relation extension.
962 : */
963 102996 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
964 88158 : flags |= EB_LOCK_TARGET;
965 :
966 210016 : while (current_size < extend_to)
967 : {
968 107020 : uint32 num_pages = lengthof(buffers);
969 : BlockNumber first_block;
970 :
971 107020 : if ((uint64) current_size + num_pages > extend_to)
972 107018 : num_pages = extend_to - current_size;
973 :
974 107020 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
975 : num_pages, extend_to,
976 : buffers, &extended_by);
977 :
978 107020 : current_size = first_block + extended_by;
979 : Assert(num_pages != 0 || current_size >= extend_to);
980 :
981 228464 : for (uint32 i = 0; i < extended_by; i++)
982 : {
983 121444 : if (first_block + i != extend_to - 1)
984 18468 : ReleaseBuffer(buffers[i]);
985 : else
986 102976 : buffer = buffers[i];
987 : }
988 : }
989 :
990 : /*
991 : * It's possible that another backend concurrently extended the relation.
992 : * In that case read the buffer.
993 : *
994 : * XXX: Should we control this via a flag?
995 : */
996 102996 : if (buffer == InvalidBuffer)
997 : {
998 : Assert(extended_by == 0);
999 20 : buffer = ReadBuffer_common(bmr.rel, BMR_GET_SMGR(bmr), bmr.relpersistence,
1000 : fork, extend_to - 1, mode, strategy);
1001 : }
1002 :
1003 102996 : return buffer;
1004 : }
1005 :
1006 : /*
1007 : * Lock and optionally zero a buffer, as part of the implementation of
1008 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1009 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1010 : */
1011 : static void
1012 634654 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1013 : {
1014 : BufferDesc *bufHdr;
1015 : bool need_to_zero;
1016 634654 : bool isLocalBuf = BufferIsLocal(buffer);
1017 :
1018 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1019 :
1020 634654 : if (already_valid)
1021 : {
1022 : /*
1023 : * If the caller already knew the buffer was valid, we can skip some
1024 : * header interaction. The caller just wants to lock the buffer.
1025 : */
1026 74000 : need_to_zero = false;
1027 : }
1028 560654 : else if (isLocalBuf)
1029 : {
1030 : /* Simple case for non-shared buffers. */
1031 48 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1032 48 : need_to_zero = StartLocalBufferIO(bufHdr, true, false);
1033 : }
1034 : else
1035 : {
1036 : /*
1037 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1038 : * concurrently. Even though we aren't doing I/O, that ensures that
1039 : * we don't zero a page that someone else has pinned. An exclusive
1040 : * content lock wouldn't be enough, because readers are allowed to
1041 : * drop the content lock after determining that a tuple is visible
1042 : * (see buffer access rules in README).
1043 : */
1044 560606 : bufHdr = GetBufferDescriptor(buffer - 1);
1045 560606 : need_to_zero = StartBufferIO(bufHdr, true, false);
1046 : }
1047 :
1048 634654 : if (need_to_zero)
1049 : {
1050 560654 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1051 :
1052 : /*
1053 : * Grab the buffer content lock before marking the page as valid, to
1054 : * make sure that no other backend sees the zeroed page before the
1055 : * caller has had a chance to initialize it.
1056 : *
1057 : * Since no-one else can be looking at the page contents yet, there is
1058 : * no difference between an exclusive lock and a cleanup-strength
1059 : * lock. (Note that we cannot use LockBuffer() or
1060 : * LockBufferForCleanup() here, because they assert that the buffer is
1061 : * already valid.)
1062 : */
1063 560654 : if (!isLocalBuf)
1064 560606 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1065 :
1066 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1067 560654 : if (isLocalBuf)
1068 48 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1069 : else
1070 560606 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1071 : }
1072 74000 : else if (!isLocalBuf)
1073 : {
1074 : /*
1075 : * The buffer is valid, so we can't zero it. The caller still expects
1076 : * the page to be locked on return.
1077 : */
1078 73960 : if (mode == RBM_ZERO_AND_LOCK)
1079 73880 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1080 : else
1081 80 : LockBufferForCleanup(buffer);
1082 : }
1083 634654 : }
1084 :
1085 : /*
1086 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1087 : * already present, or false if more work is required to either read it in or
1088 : * zero it.
1089 : */
1090 : static pg_attribute_always_inline Buffer
1091 124392898 : PinBufferForBlock(Relation rel,
1092 : SMgrRelation smgr,
1093 : char persistence,
1094 : ForkNumber forkNum,
1095 : BlockNumber blockNum,
1096 : BufferAccessStrategy strategy,
1097 : bool *foundPtr)
1098 : {
1099 : BufferDesc *bufHdr;
1100 : IOContext io_context;
1101 : IOObject io_object;
1102 :
1103 : Assert(blockNum != P_NEW);
1104 :
1105 : /* Persistence should be set before */
1106 : Assert((persistence == RELPERSISTENCE_TEMP ||
1107 : persistence == RELPERSISTENCE_PERMANENT ||
1108 : persistence == RELPERSISTENCE_UNLOGGED));
1109 :
1110 124392898 : if (persistence == RELPERSISTENCE_TEMP)
1111 : {
1112 2554270 : io_context = IOCONTEXT_NORMAL;
1113 2554270 : io_object = IOOBJECT_TEMP_RELATION;
1114 : }
1115 : else
1116 : {
1117 121838628 : io_context = IOContextForStrategy(strategy);
1118 121838628 : io_object = IOOBJECT_RELATION;
1119 : }
1120 :
1121 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1122 : smgr->smgr_rlocator.locator.spcOid,
1123 : smgr->smgr_rlocator.locator.dbOid,
1124 : smgr->smgr_rlocator.locator.relNumber,
1125 : smgr->smgr_rlocator.backend);
1126 :
1127 124392898 : if (persistence == RELPERSISTENCE_TEMP)
1128 : {
1129 2554270 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1130 2554258 : if (*foundPtr)
1131 2537476 : pgBufferUsage.local_blks_hit++;
1132 : }
1133 : else
1134 : {
1135 121838628 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1136 : strategy, foundPtr, io_context);
1137 121838628 : if (*foundPtr)
1138 118432264 : pgBufferUsage.shared_blks_hit++;
1139 : }
1140 124392886 : if (rel)
1141 : {
1142 : /*
1143 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1144 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1145 : * zeroed instead), the per-relation stats always count them.
1146 : */
1147 112551422 : pgstat_count_buffer_read(rel);
1148 112551422 : if (*foundPtr)
1149 110054548 : pgstat_count_buffer_hit(rel);
1150 : }
1151 124392886 : if (*foundPtr)
1152 : {
1153 120969740 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1154 120969740 : if (VacuumCostActive)
1155 6209540 : VacuumCostBalance += VacuumCostPageHit;
1156 :
1157 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forkNum, blockNum,
1158 : smgr->smgr_rlocator.locator.spcOid,
1159 : smgr->smgr_rlocator.locator.dbOid,
1160 : smgr->smgr_rlocator.locator.relNumber,
1161 : smgr->smgr_rlocator.backend,
1162 : true);
1163 : }
1164 :
1165 124392886 : return BufferDescriptorGetBuffer(bufHdr);
1166 : }
1167 :
1168 : /*
1169 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1170 : *
1171 : * smgr is required, rel is optional unless using P_NEW.
1172 : */
1173 : static pg_attribute_always_inline Buffer
1174 116105420 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1175 : ForkNumber forkNum,
1176 : BlockNumber blockNum, ReadBufferMode mode,
1177 : BufferAccessStrategy strategy)
1178 : {
1179 : ReadBuffersOperation operation;
1180 : Buffer buffer;
1181 : int flags;
1182 : char persistence;
1183 :
1184 : /*
1185 : * Backward compatibility path, most code should use ExtendBufferedRel()
1186 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1187 : * scales a lot better.
1188 : */
1189 116105420 : if (unlikely(blockNum == P_NEW))
1190 : {
1191 522 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1192 :
1193 : /*
1194 : * Since no-one else can be looking at the page contents yet, there is
1195 : * no difference between an exclusive lock and a cleanup-strength
1196 : * lock.
1197 : */
1198 522 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1199 0 : flags |= EB_LOCK_FIRST;
1200 :
1201 522 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1202 : }
1203 :
1204 116104898 : if (rel)
1205 104731800 : persistence = rel->rd_rel->relpersistence;
1206 : else
1207 11373098 : persistence = smgr_persistence;
1208 :
1209 116104898 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1210 : mode == RBM_ZERO_AND_LOCK))
1211 : {
1212 : bool found;
1213 :
1214 634654 : buffer = PinBufferForBlock(rel, smgr, persistence,
1215 : forkNum, blockNum, strategy, &found);
1216 634654 : ZeroAndLockBuffer(buffer, mode, found);
1217 634654 : return buffer;
1218 : }
1219 :
1220 : /*
1221 : * Signal that we are going to immediately wait. If we're immediately
1222 : * waiting, there is no benefit in actually executing the IO
1223 : * asynchronously, it would just add dispatch overhead.
1224 : */
1225 115470244 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1226 115470244 : if (mode == RBM_ZERO_ON_ERROR)
1227 3092414 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1228 115470244 : operation.smgr = smgr;
1229 115470244 : operation.rel = rel;
1230 115470244 : operation.persistence = persistence;
1231 115470244 : operation.forknum = forkNum;
1232 115470244 : operation.strategy = strategy;
1233 115470244 : if (StartReadBuffer(&operation,
1234 : &buffer,
1235 : blockNum,
1236 : flags))
1237 1408340 : WaitReadBuffers(&operation);
1238 :
1239 115470198 : return buffer;
1240 : }
1241 :
1242 : static pg_attribute_always_inline bool
1243 123423678 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1244 : Buffer *buffers,
1245 : BlockNumber blockNum,
1246 : int *nblocks,
1247 : int flags,
1248 : bool allow_forwarding)
1249 : {
1250 123423678 : int actual_nblocks = *nblocks;
1251 123423678 : int maxcombine = 0;
1252 : bool did_start_io;
1253 :
1254 : Assert(*nblocks == 1 || allow_forwarding);
1255 : Assert(*nblocks > 0);
1256 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1257 :
1258 126286170 : for (int i = 0; i < actual_nblocks; ++i)
1259 : {
1260 : bool found;
1261 :
1262 123761070 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1263 2826 : {
1264 : BufferDesc *bufHdr;
1265 :
1266 : /*
1267 : * This is a buffer that was pinned by an earlier call to
1268 : * StartReadBuffers(), but couldn't be handled in one operation at
1269 : * that time. The operation was split, and the caller has passed
1270 : * an already pinned buffer back to us to handle the rest of the
1271 : * operation. It must continue at the expected block number.
1272 : */
1273 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1274 :
1275 : /*
1276 : * It might be an already valid buffer (a hit) that followed the
1277 : * final contiguous block of an earlier I/O (a miss) marking the
1278 : * end of it, or a buffer that some other backend has since made
1279 : * valid by performing the I/O for us, in which case we can handle
1280 : * it as a hit now. It is safe to check for a BM_VALID flag with
1281 : * a relaxed load, because we got a fresh view of it while pinning
1282 : * it in the previous call.
1283 : *
1284 : * On the other hand if we don't see BM_VALID yet, it must be an
1285 : * I/O that was split by the previous call and we need to try to
1286 : * start a new I/O from this block. We're also racing against any
1287 : * other backend that might start the I/O or even manage to mark
1288 : * it BM_VALID after this check, but StartBufferIO() will handle
1289 : * those cases.
1290 : */
1291 2826 : if (BufferIsLocal(buffers[i]))
1292 4 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1293 : else
1294 2822 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1295 : Assert(pg_atomic_read_u32(&bufHdr->state) & BM_TAG_VALID);
1296 2826 : found = pg_atomic_read_u32(&bufHdr->state) & BM_VALID;
1297 : }
1298 : else
1299 : {
1300 123758232 : buffers[i] = PinBufferForBlock(operation->rel,
1301 : operation->smgr,
1302 123758244 : operation->persistence,
1303 : operation->forknum,
1304 : blockNum + i,
1305 : operation->strategy,
1306 : &found);
1307 : }
1308 :
1309 123761058 : if (found)
1310 : {
1311 : /*
1312 : * We have a hit. If it's the first block in the requested range,
1313 : * we can return it immediately and report that WaitReadBuffers()
1314 : * does not need to be called. If the initial value of *nblocks
1315 : * was larger, the caller will have to call again for the rest.
1316 : */
1317 120898566 : if (i == 0)
1318 : {
1319 120895736 : *nblocks = 1;
1320 :
1321 : #ifdef USE_ASSERT_CHECKING
1322 :
1323 : /*
1324 : * Initialize enough of ReadBuffersOperation to make
1325 : * CheckReadBuffersOperation() work. Outside of assertions
1326 : * that's not necessary when no IO is issued.
1327 : */
1328 : operation->buffers = buffers;
1329 : operation->blocknum = blockNum;
1330 : operation->nblocks = 1;
1331 : operation->nblocks_done = 1;
1332 : CheckReadBuffersOperation(operation, true);
1333 : #endif
1334 120895736 : return false;
1335 : }
1336 :
1337 : /*
1338 : * Otherwise we already have an I/O to perform, but this block
1339 : * can't be included as it is already valid. Split the I/O here.
1340 : * There may or may not be more blocks requiring I/O after this
1341 : * one, we haven't checked, but they can't be contiguous with this
1342 : * one in the way. We'll leave this buffer pinned, forwarding it
1343 : * to the next call, avoiding the need to unpin it here and re-pin
1344 : * it in the next call.
1345 : */
1346 2830 : actual_nblocks = i;
1347 2830 : break;
1348 : }
1349 : else
1350 : {
1351 : /*
1352 : * Check how many blocks we can cover with the same IO. The smgr
1353 : * implementation might e.g. be limited due to a segment boundary.
1354 : */
1355 2862492 : if (i == 0 && actual_nblocks > 1)
1356 : {
1357 67730 : maxcombine = smgrmaxcombine(operation->smgr,
1358 : operation->forknum,
1359 : blockNum);
1360 67730 : if (unlikely(maxcombine < actual_nblocks))
1361 : {
1362 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1363 : blockNum, actual_nblocks, maxcombine);
1364 0 : actual_nblocks = maxcombine;
1365 : }
1366 : }
1367 : }
1368 : }
1369 2527930 : *nblocks = actual_nblocks;
1370 :
1371 : /* Populate information needed for I/O. */
1372 2527930 : operation->buffers = buffers;
1373 2527930 : operation->blocknum = blockNum;
1374 2527930 : operation->flags = flags;
1375 2527930 : operation->nblocks = actual_nblocks;
1376 2527930 : operation->nblocks_done = 0;
1377 2527930 : pgaio_wref_clear(&operation->io_wref);
1378 :
1379 : /*
1380 : * When using AIO, start the IO in the background. If not, issue prefetch
1381 : * requests if desired by the caller.
1382 : *
1383 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1384 : * de-risk the introduction of AIO somewhat. It's a large architectural
1385 : * change, with lots of chances for unanticipated performance effects.
1386 : *
1387 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1388 : * asynchronously, but without the check here we'd execute IO earlier than
1389 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1390 : */
1391 2527930 : if (io_method != IOMETHOD_SYNC)
1392 : {
1393 : /*
1394 : * Try to start IO asynchronously. It's possible that no IO needs to
1395 : * be started, if another backend already performed the IO.
1396 : *
1397 : * Note that if an IO is started, it might not cover the entire
1398 : * requested range, e.g. because an intermediary block has been read
1399 : * in by another backend. In that case any "trailing" buffers we
1400 : * already pinned above will be "forwarded" by read_stream.c to the
1401 : * next call to StartReadBuffers().
1402 : *
1403 : * This is signalled to the caller by decrementing *nblocks *and*
1404 : * reducing operation->nblocks. The latter is done here, but not below
1405 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1406 : * overall read size anymore, we need to retry until done in its
1407 : * entirety or until failed.
1408 : */
1409 2525776 : did_start_io = AsyncReadBuffers(operation, nblocks);
1410 :
1411 2525746 : operation->nblocks = *nblocks;
1412 : }
1413 : else
1414 : {
1415 2154 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1416 :
1417 2154 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1418 : {
1419 : /*
1420 : * In theory we should only do this if PinBufferForBlock() had to
1421 : * allocate new buffers above. That way, if two calls to
1422 : * StartReadBuffers() were made for the same blocks before
1423 : * WaitReadBuffers(), only the first would issue the advice.
1424 : * That'd be a better simulation of true asynchronous I/O, which
1425 : * would only start the I/O once, but isn't done here for
1426 : * simplicity.
1427 : */
1428 4 : smgrprefetch(operation->smgr,
1429 : operation->forknum,
1430 : blockNum,
1431 : actual_nblocks);
1432 : }
1433 :
1434 : /*
1435 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1436 : * will initiate the necessary IO.
1437 : */
1438 2154 : did_start_io = true;
1439 : }
1440 :
1441 2527900 : CheckReadBuffersOperation(operation, !did_start_io);
1442 :
1443 2527900 : return did_start_io;
1444 : }
1445 :
1446 : /*
1447 : * Begin reading a range of blocks beginning at blockNum and extending for
1448 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1449 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1450 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1451 : * and is now being continued. On return, *nblocks holds the number of blocks
1452 : * accepted by this operation. If it is less than the original number then
1453 : * this operation has been split, but buffer elements up to the original
1454 : * requested size may hold forwarded buffers to be used for a continuing
1455 : * operation. The caller must either start a new I/O beginning at the block
1456 : * immediately following the blocks accepted by this call and pass those
1457 : * buffers back in, or release them if it chooses not to. It shouldn't make
1458 : * any other use of or assumptions about forwarded buffers.
1459 : *
1460 : * If false is returned, no I/O is necessary and the buffers covered by
1461 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1462 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1463 : * operation object before the buffers covered by *nblocks on exit can be
1464 : * accessed. Along with the operation object, the caller-supplied array of
1465 : * buffers must remain valid until WaitReadBuffers() is called, and any
1466 : * forwarded buffers must also be preserved for a continuing call unless
1467 : * they are explicitly released.
1468 : */
1469 : bool
1470 3912318 : StartReadBuffers(ReadBuffersOperation *operation,
1471 : Buffer *buffers,
1472 : BlockNumber blockNum,
1473 : int *nblocks,
1474 : int flags)
1475 : {
1476 3912318 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1477 : true /* expect forwarded buffers */ );
1478 : }
1479 :
1480 : /*
1481 : * Single block version of the StartReadBuffers(). This might save a few
1482 : * instructions when called from another translation unit, because it is
1483 : * specialized for nblocks == 1.
1484 : *
1485 : * This version does not support "forwarded" buffers: they cannot be created
1486 : * by reading only one block and *buffer is ignored on entry.
1487 : */
1488 : bool
1489 119511360 : StartReadBuffer(ReadBuffersOperation *operation,
1490 : Buffer *buffer,
1491 : BlockNumber blocknum,
1492 : int flags)
1493 : {
1494 119511360 : int nblocks = 1;
1495 : bool result;
1496 :
1497 119511360 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1498 : false /* single block, no forwarding */ );
1499 : Assert(nblocks == 1); /* single block can't be short */
1500 :
1501 119511330 : return result;
1502 : }
1503 :
1504 : /*
1505 : * Perform sanity checks on the ReadBuffersOperation.
1506 : */
1507 : static void
1508 7554684 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1509 : {
1510 : #ifdef USE_ASSERT_CHECKING
1511 : Assert(operation->nblocks_done <= operation->nblocks);
1512 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1513 :
1514 : for (int i = 0; i < operation->nblocks; i++)
1515 : {
1516 : Buffer buffer = operation->buffers[i];
1517 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1518 : GetLocalBufferDescriptor(-buffer - 1) :
1519 : GetBufferDescriptor(buffer - 1);
1520 :
1521 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1522 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_TAG_VALID);
1523 :
1524 : if (i < operation->nblocks_done)
1525 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_VALID);
1526 : }
1527 : #endif
1528 7554684 : }
1529 :
1530 : /* helper for ReadBuffersCanStartIO(), to avoid repetition */
1531 : static inline bool
1532 2862528 : ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
1533 : {
1534 2862528 : if (BufferIsLocal(buffer))
1535 16734 : return StartLocalBufferIO(GetLocalBufferDescriptor(-buffer - 1),
1536 : true, nowait);
1537 : else
1538 2845794 : return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
1539 : }
1540 :
1541 : /*
1542 : * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
1543 : */
1544 : static inline bool
1545 2862528 : ReadBuffersCanStartIO(Buffer buffer, bool nowait)
1546 : {
1547 : /*
1548 : * If this backend currently has staged IO, we need to submit the pending
1549 : * IO before waiting for the right to issue IO, to avoid the potential for
1550 : * deadlocks (and, more commonly, unnecessary delays for other backends).
1551 : */
1552 2862528 : if (!nowait && pgaio_have_staged())
1553 : {
1554 1154 : if (ReadBuffersCanStartIOOnce(buffer, true))
1555 1154 : return true;
1556 :
1557 : /*
1558 : * Unfortunately StartBufferIO() returning false doesn't allow to
1559 : * distinguish between the buffer already being valid and IO already
1560 : * being in progress. Since IO already being in progress is quite
1561 : * rare, this approach seems fine.
1562 : */
1563 0 : pgaio_submit_staged();
1564 : }
1565 :
1566 2861374 : return ReadBuffersCanStartIOOnce(buffer, nowait);
1567 : }
1568 :
1569 : /*
1570 : * Helper for WaitReadBuffers() that processes the results of a readv
1571 : * operation, raising an error if necessary.
1572 : */
1573 : static void
1574 2512354 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1575 : {
1576 2512354 : PgAioReturn *aio_ret = &operation->io_return;
1577 2512354 : PgAioResultStatus rs = aio_ret->result.status;
1578 2512354 : int newly_read_blocks = 0;
1579 :
1580 : Assert(pgaio_wref_valid(&operation->io_wref));
1581 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1582 :
1583 : /*
1584 : * SMGR reports the number of blocks successfully read as the result of
1585 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1586 : */
1587 :
1588 2512354 : if (likely(rs != PGAIO_RS_ERROR))
1589 2512296 : newly_read_blocks = aio_ret->result.result;
1590 :
1591 2512354 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1592 90 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1593 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1594 2512264 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1595 : {
1596 : /*
1597 : * We'll retry, so we just emit a debug message to the server log (or
1598 : * not even that in prod scenarios).
1599 : */
1600 20 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1601 20 : elog(DEBUG3, "partial read, will retry");
1602 : }
1603 :
1604 : Assert(newly_read_blocks > 0);
1605 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1606 :
1607 2512296 : operation->nblocks_done += newly_read_blocks;
1608 :
1609 : Assert(operation->nblocks_done <= operation->nblocks);
1610 2512296 : }
1611 :
1612 : void
1613 2512334 : WaitReadBuffers(ReadBuffersOperation *operation)
1614 : {
1615 2512334 : PgAioReturn *aio_ret = &operation->io_return;
1616 : IOContext io_context;
1617 : IOObject io_object;
1618 :
1619 2512334 : if (operation->persistence == RELPERSISTENCE_TEMP)
1620 : {
1621 2980 : io_context = IOCONTEXT_NORMAL;
1622 2980 : io_object = IOOBJECT_TEMP_RELATION;
1623 : }
1624 : else
1625 : {
1626 2509354 : io_context = IOContextForStrategy(operation->strategy);
1627 2509354 : io_object = IOOBJECT_RELATION;
1628 : }
1629 :
1630 : /*
1631 : * If we get here without an IO operation having been issued, the
1632 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1633 : * caller should not have called WaitReadBuffers().
1634 : *
1635 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1636 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1637 : * of the retry logic below, no extra code is required.
1638 : *
1639 : * This path is expected to eventually go away.
1640 : */
1641 2512334 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1642 0 : elog(ERROR, "waiting for read operation that didn't read");
1643 :
1644 : /*
1645 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1646 : * done. We may need multiple retries, not just because we could get
1647 : * multiple partial reads, but also because some of the remaining
1648 : * to-be-read buffers may have been read in by other backends, limiting
1649 : * the IO size.
1650 : */
1651 : while (true)
1652 2174 : {
1653 : int ignored_nblocks_progress;
1654 :
1655 2514508 : CheckReadBuffersOperation(operation, false);
1656 :
1657 : /*
1658 : * If there is an IO associated with the operation, we may need to
1659 : * wait for it.
1660 : */
1661 2514508 : if (pgaio_wref_valid(&operation->io_wref))
1662 : {
1663 : /*
1664 : * Track the time spent waiting for the IO to complete. As
1665 : * tracking a wait even if we don't actually need to wait
1666 : *
1667 : * a) is not cheap, due to the timestamping overhead
1668 : *
1669 : * b) reports some time as waiting, even if we never waited
1670 : *
1671 : * we first check if we already know the IO is complete.
1672 : */
1673 2512354 : if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
1674 1087796 : !pgaio_wref_check_done(&operation->io_wref))
1675 : {
1676 417460 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1677 :
1678 417460 : pgaio_wref_wait(&operation->io_wref);
1679 :
1680 : /*
1681 : * The IO operation itself was already counted earlier, in
1682 : * AsyncReadBuffers(), this just accounts for the wait time.
1683 : */
1684 417460 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1685 : io_start, 0, 0);
1686 : }
1687 : else
1688 : {
1689 : Assert(pgaio_wref_check_done(&operation->io_wref));
1690 : }
1691 :
1692 : /*
1693 : * We now are sure the IO completed. Check the results. This
1694 : * includes reporting on errors if there were any.
1695 : */
1696 2512354 : ProcessReadBuffersResult(operation);
1697 : }
1698 :
1699 : /*
1700 : * Most of the time, the one IO we already started, will read in
1701 : * everything. But we need to deal with partial reads and buffers not
1702 : * needing IO anymore.
1703 : */
1704 2514450 : if (operation->nblocks_done == operation->nblocks)
1705 2512276 : break;
1706 :
1707 2174 : CHECK_FOR_INTERRUPTS();
1708 :
1709 : /*
1710 : * This may only complete the IO partially, either because some
1711 : * buffers were already valid, or because of a partial read.
1712 : *
1713 : * NB: In contrast to after the AsyncReadBuffers() call in
1714 : * StartReadBuffers(), we do *not* reduce
1715 : * ReadBuffersOperation->nblocks here, callers expect the full
1716 : * operation to be completed at this point (as more operations may
1717 : * have been queued).
1718 : */
1719 2174 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1720 : }
1721 :
1722 2512276 : CheckReadBuffersOperation(operation, true);
1723 :
1724 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1725 2512276 : }
1726 :
1727 : /*
1728 : * Initiate IO for the ReadBuffersOperation
1729 : *
1730 : * This function only starts a single IO at a time. The size of the IO may be
1731 : * limited to below the to-be-read blocks, if one of the buffers has
1732 : * concurrently been read in. If the first to-be-read buffer is already valid,
1733 : * no IO will be issued.
1734 : *
1735 : * To support retries after partial reads, the first operation->nblocks_done
1736 : * buffers are skipped.
1737 : *
1738 : * On return *nblocks_progress is updated to reflect the number of buffers
1739 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1740 : * set to 1 and operation->nblocks_done is incremented.
1741 : *
1742 : * Returns true if IO was initiated, false if no IO was necessary.
1743 : */
1744 : static bool
1745 2527950 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1746 : {
1747 2527950 : Buffer *buffers = &operation->buffers[0];
1748 2527950 : int flags = operation->flags;
1749 2527950 : BlockNumber blocknum = operation->blocknum;
1750 2527950 : ForkNumber forknum = operation->forknum;
1751 2527950 : char persistence = operation->persistence;
1752 2527950 : int16 nblocks_done = operation->nblocks_done;
1753 2527950 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1754 2527950 : int io_buffers_len = 0;
1755 : PgAioHandle *ioh;
1756 2527950 : uint32 ioh_flags = 0;
1757 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1758 : IOContext io_context;
1759 : IOObject io_object;
1760 : bool did_start_io;
1761 :
1762 : /*
1763 : * When this IO is executed synchronously, either because the caller will
1764 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1765 : * the AIO subsystem needs to know.
1766 : */
1767 2527950 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1768 1419296 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1769 :
1770 2527950 : if (persistence == RELPERSISTENCE_TEMP)
1771 : {
1772 3568 : io_context = IOCONTEXT_NORMAL;
1773 3568 : io_object = IOOBJECT_TEMP_RELATION;
1774 3568 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1775 : }
1776 : else
1777 : {
1778 2524382 : io_context = IOContextForStrategy(operation->strategy);
1779 2524382 : io_object = IOOBJECT_RELATION;
1780 : }
1781 :
1782 : /*
1783 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1784 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1785 : * set globally, but on a per-session basis. The completion callback,
1786 : * which may be run in other processes, e.g. in IO workers, may have a
1787 : * different value of the zero_damaged_pages GUC.
1788 : *
1789 : * XXX: We probably should eventually use a different flag for
1790 : * zero_damaged_pages, so we can report different log levels / error codes
1791 : * for zero_damaged_pages and ZERO_ON_ERROR.
1792 : */
1793 2527950 : if (zero_damaged_pages)
1794 32 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1795 :
1796 : /*
1797 : * For the same reason as with zero_damaged_pages we need to use this
1798 : * backend's ignore_checksum_failure value.
1799 : */
1800 2527950 : if (ignore_checksum_failure)
1801 16 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1802 :
1803 :
1804 : /*
1805 : * To be allowed to report stats in the local completion callback we need
1806 : * to prepare to report stats now. This ensures we can safely report the
1807 : * checksum failure even in a critical section.
1808 : */
1809 2527950 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
1810 :
1811 : /*
1812 : * Get IO handle before ReadBuffersCanStartIO(), as pgaio_io_acquire()
1813 : * might block, which we don't want after setting IO_IN_PROGRESS.
1814 : *
1815 : * If we need to wait for IO before we can get a handle, submit
1816 : * already-staged IO first, so that other backends don't need to wait.
1817 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
1818 : * wait for already submitted IO, which doesn't require additional locks,
1819 : * but it could still cause undesirable waits.
1820 : *
1821 : * A secondary benefit is that this would allow us to measure the time in
1822 : * pgaio_io_acquire() without causing undue timer overhead in the common,
1823 : * non-blocking, case. However, currently the pgstats infrastructure
1824 : * doesn't really allow that, as it a) asserts that an operation can't
1825 : * have time without operations b) doesn't have an API to report
1826 : * "accumulated" time.
1827 : */
1828 2527950 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
1829 2527950 : if (unlikely(!ioh))
1830 : {
1831 6280 : pgaio_submit_staged();
1832 :
1833 6280 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
1834 : }
1835 :
1836 : /*
1837 : * Check if we can start IO on the first to-be-read buffer.
1838 : *
1839 : * If an I/O is already in progress in another backend, we want to wait
1840 : * for the outcome: either done, or something went wrong and we will
1841 : * retry.
1842 : */
1843 2527950 : if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
1844 : {
1845 : /*
1846 : * Someone else has already completed this block, we're done.
1847 : *
1848 : * When IO is necessary, ->nblocks_done is updated in
1849 : * ProcessReadBuffersResult(), but that is not called if no IO is
1850 : * necessary. Thus update here.
1851 : */
1852 14972 : operation->nblocks_done += 1;
1853 14972 : *nblocks_progress = 1;
1854 :
1855 14972 : pgaio_io_release(ioh);
1856 14972 : pgaio_wref_clear(&operation->io_wref);
1857 14972 : did_start_io = false;
1858 :
1859 : /*
1860 : * Report and track this as a 'hit' for this backend, even though it
1861 : * must have started out as a miss in PinBufferForBlock(). The other
1862 : * backend will track this as a 'read'.
1863 : */
1864 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum, blocknum + operation->nblocks_done,
1865 : operation->smgr->smgr_rlocator.locator.spcOid,
1866 : operation->smgr->smgr_rlocator.locator.dbOid,
1867 : operation->smgr->smgr_rlocator.locator.relNumber,
1868 : operation->smgr->smgr_rlocator.backend,
1869 : true);
1870 :
1871 14972 : if (persistence == RELPERSISTENCE_TEMP)
1872 0 : pgBufferUsage.local_blks_hit += 1;
1873 : else
1874 14972 : pgBufferUsage.shared_blks_hit += 1;
1875 :
1876 14972 : if (operation->rel)
1877 14972 : pgstat_count_buffer_hit(operation->rel);
1878 :
1879 14972 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1880 :
1881 14972 : if (VacuumCostActive)
1882 12 : VacuumCostBalance += VacuumCostPageHit;
1883 : }
1884 : else
1885 : {
1886 : instr_time io_start;
1887 :
1888 : /* We found a buffer that we need to read in. */
1889 : Assert(io_buffers[0] == buffers[nblocks_done]);
1890 2512978 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
1891 2512978 : io_buffers_len = 1;
1892 :
1893 : /*
1894 : * How many neighboring-on-disk blocks can we scatter-read into other
1895 : * buffers at the same time? In this case we don't wait if we see an
1896 : * I/O already in progress. We already set BM_IO_IN_PROGRESS for the
1897 : * head block, so we should get on with that I/O as soon as possible.
1898 : */
1899 2847556 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
1900 : {
1901 334578 : if (!ReadBuffersCanStartIO(buffers[i], true))
1902 0 : break;
1903 : /* Must be consecutive block numbers. */
1904 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
1905 : BufferGetBlockNumber(buffers[i]) - 1);
1906 : Assert(io_buffers[io_buffers_len] == buffers[i]);
1907 :
1908 334578 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
1909 : }
1910 :
1911 : /* get a reference to wait for in WaitReadBuffers() */
1912 2512978 : pgaio_io_get_wref(ioh, &operation->io_wref);
1913 :
1914 : /* provide the list of buffers to the completion callbacks */
1915 2512978 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
1916 :
1917 2512978 : pgaio_io_register_callbacks(ioh,
1918 : persistence == RELPERSISTENCE_TEMP ?
1919 : PGAIO_HCB_LOCAL_BUFFER_READV :
1920 : PGAIO_HCB_SHARED_BUFFER_READV,
1921 : flags);
1922 :
1923 2512978 : pgaio_io_set_flag(ioh, ioh_flags);
1924 :
1925 : /* ---
1926 : * Even though we're trying to issue IO asynchronously, track the time
1927 : * in smgrstartreadv():
1928 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
1929 : * immediately
1930 : * - the io method might not support the IO (e.g. worker IO for a temp
1931 : * table)
1932 : * ---
1933 : */
1934 2512978 : io_start = pgstat_prepare_io_time(track_io_timing);
1935 2512978 : smgrstartreadv(ioh, operation->smgr, forknum,
1936 : blocknum + nblocks_done,
1937 : io_pages, io_buffers_len);
1938 2512948 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1939 2512948 : io_start, 1, io_buffers_len * BLCKSZ);
1940 :
1941 2512948 : if (persistence == RELPERSISTENCE_TEMP)
1942 3568 : pgBufferUsage.local_blks_read += io_buffers_len;
1943 : else
1944 2509380 : pgBufferUsage.shared_blks_read += io_buffers_len;
1945 :
1946 : /*
1947 : * Track vacuum cost when issuing IO, not after waiting for it.
1948 : * Otherwise we could end up issuing a lot of IO in a short timespan,
1949 : * despite a low cost limit.
1950 : */
1951 2512948 : if (VacuumCostActive)
1952 44124 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
1953 :
1954 2512948 : *nblocks_progress = io_buffers_len;
1955 2512948 : did_start_io = true;
1956 : }
1957 :
1958 2527920 : return did_start_io;
1959 : }
1960 :
1961 : /*
1962 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
1963 : * buffer. If no buffer exists already, selects a replacement victim and
1964 : * evicts the old page, but does NOT read in new page.
1965 : *
1966 : * "strategy" can be a buffer replacement strategy object, or NULL for
1967 : * the default strategy. The selected buffer's usage_count is advanced when
1968 : * using the default strategy, but otherwise possibly not (see PinBuffer).
1969 : *
1970 : * The returned buffer is pinned and is already marked as holding the
1971 : * desired page. If it already did have the desired page, *foundPtr is
1972 : * set true. Otherwise, *foundPtr is set false.
1973 : *
1974 : * io_context is passed as an output parameter to avoid calling
1975 : * IOContextForStrategy() when there is a shared buffers hit and no IO
1976 : * statistics need be captured.
1977 : *
1978 : * No locks are held either at entry or exit.
1979 : */
1980 : static pg_attribute_always_inline BufferDesc *
1981 121838792 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
1982 : BlockNumber blockNum,
1983 : BufferAccessStrategy strategy,
1984 : bool *foundPtr, IOContext io_context)
1985 : {
1986 : BufferTag newTag; /* identity of requested block */
1987 : uint32 newHash; /* hash value for newTag */
1988 : LWLock *newPartitionLock; /* buffer partition lock for it */
1989 : int existing_buf_id;
1990 : Buffer victim_buffer;
1991 : BufferDesc *victim_buf_hdr;
1992 : uint32 victim_buf_state;
1993 121838792 : uint32 set_bits = 0;
1994 :
1995 : /* Make sure we will have room to remember the buffer pin */
1996 121838792 : ResourceOwnerEnlarge(CurrentResourceOwner);
1997 121838792 : ReservePrivateRefCountEntry();
1998 :
1999 : /* create a tag so we can lookup the buffer */
2000 121838792 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2001 :
2002 : /* determine its hash code and partition lock ID */
2003 121838792 : newHash = BufTableHashCode(&newTag);
2004 121838792 : newPartitionLock = BufMappingPartitionLock(newHash);
2005 :
2006 : /* see if the block is in the buffer pool already */
2007 121838792 : LWLockAcquire(newPartitionLock, LW_SHARED);
2008 121838792 : existing_buf_id = BufTableLookup(&newTag, newHash);
2009 121838792 : if (existing_buf_id >= 0)
2010 : {
2011 : BufferDesc *buf;
2012 : bool valid;
2013 :
2014 : /*
2015 : * Found it. Now, pin the buffer so no one can steal it from the
2016 : * buffer pool, and check to see if the correct data has been loaded
2017 : * into the buffer.
2018 : */
2019 118444678 : buf = GetBufferDescriptor(existing_buf_id);
2020 :
2021 118444678 : valid = PinBuffer(buf, strategy, false);
2022 :
2023 : /* Can release the mapping lock as soon as we've pinned it */
2024 118444678 : LWLockRelease(newPartitionLock);
2025 :
2026 118444678 : *foundPtr = true;
2027 :
2028 118444678 : if (!valid)
2029 : {
2030 : /*
2031 : * We can only get here if (a) someone else is still reading in
2032 : * the page, (b) a previous read attempt failed, or (c) someone
2033 : * called StartReadBuffers() but not yet WaitReadBuffers().
2034 : */
2035 12954 : *foundPtr = false;
2036 : }
2037 :
2038 118444678 : return buf;
2039 : }
2040 :
2041 : /*
2042 : * Didn't find it in the buffer pool. We'll have to initialize a new
2043 : * buffer. Remember to unlock the mapping lock while doing the work.
2044 : */
2045 3394114 : LWLockRelease(newPartitionLock);
2046 :
2047 : /*
2048 : * Acquire a victim buffer. Somebody else might try to do the same, we
2049 : * don't hold any conflicting locks. If so we'll have to undo our work
2050 : * later.
2051 : */
2052 3394114 : victim_buffer = GetVictimBuffer(strategy, io_context);
2053 3394114 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2054 :
2055 : /*
2056 : * Try to make a hashtable entry for the buffer under its new tag. If
2057 : * somebody else inserted another buffer for the tag, we'll release the
2058 : * victim buffer we acquired and use the already inserted one.
2059 : */
2060 3394114 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2061 3394114 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2062 3394114 : if (existing_buf_id >= 0)
2063 : {
2064 : BufferDesc *existing_buf_hdr;
2065 : bool valid;
2066 :
2067 : /*
2068 : * Got a collision. Someone has already done what we were about to do.
2069 : * We'll just handle this as if it were found in the buffer pool in
2070 : * the first place. First, give up the buffer we were planning to
2071 : * use.
2072 : *
2073 : * We could do this after releasing the partition lock, but then we'd
2074 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2075 : * before acquiring the lock, for the rare case of such a collision.
2076 : */
2077 2818 : UnpinBuffer(victim_buf_hdr);
2078 :
2079 : /* remaining code should match code at top of routine */
2080 :
2081 2818 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2082 :
2083 2818 : valid = PinBuffer(existing_buf_hdr, strategy, false);
2084 :
2085 : /* Can release the mapping lock as soon as we've pinned it */
2086 2818 : LWLockRelease(newPartitionLock);
2087 :
2088 2818 : *foundPtr = true;
2089 :
2090 2818 : if (!valid)
2091 : {
2092 : /*
2093 : * We can only get here if (a) someone else is still reading in
2094 : * the page, (b) a previous read attempt failed, or (c) someone
2095 : * called StartReadBuffers() but not yet WaitReadBuffers().
2096 : */
2097 2114 : *foundPtr = false;
2098 : }
2099 :
2100 2818 : return existing_buf_hdr;
2101 : }
2102 :
2103 : /*
2104 : * Need to lock the buffer header too in order to change its tag.
2105 : */
2106 3391296 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2107 :
2108 : /* some sanity checks while we hold the buffer header lock */
2109 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2110 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2111 :
2112 3391296 : victim_buf_hdr->tag = newTag;
2113 :
2114 : /*
2115 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2116 : * checkpoint. Unlogged buffers only need to be written at shutdown
2117 : * checkpoints, except for their "init" forks, which need to be treated
2118 : * just like permanent relations.
2119 : */
2120 3391296 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2121 3391296 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2122 3390614 : set_bits |= BM_PERMANENT;
2123 :
2124 3391296 : UnlockBufHdrExt(victim_buf_hdr, victim_buf_state,
2125 : set_bits, 0, 0);
2126 :
2127 3391296 : LWLockRelease(newPartitionLock);
2128 :
2129 : /*
2130 : * Buffer contents are currently invalid.
2131 : */
2132 3391296 : *foundPtr = false;
2133 :
2134 3391296 : return victim_buf_hdr;
2135 : }
2136 :
2137 : /*
2138 : * InvalidateBuffer -- mark a shared buffer invalid.
2139 : *
2140 : * The buffer header spinlock must be held at entry. We drop it before
2141 : * returning. (This is sane because the caller must have locked the
2142 : * buffer in order to be sure it should be dropped.)
2143 : *
2144 : * This is used only in contexts such as dropping a relation. We assume
2145 : * that no other backend could possibly be interested in using the page,
2146 : * so the only reason the buffer might be pinned is if someone else is
2147 : * trying to write it out. We have to let them finish before we can
2148 : * reclaim the buffer.
2149 : *
2150 : * The buffer could get reclaimed by someone else while we are waiting
2151 : * to acquire the necessary locks; if so, don't mess it up.
2152 : */
2153 : static void
2154 211334 : InvalidateBuffer(BufferDesc *buf)
2155 : {
2156 : BufferTag oldTag;
2157 : uint32 oldHash; /* hash value for oldTag */
2158 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2159 : uint32 oldFlags;
2160 : uint32 buf_state;
2161 :
2162 : /* Save the original buffer tag before dropping the spinlock */
2163 211334 : oldTag = buf->tag;
2164 :
2165 211334 : UnlockBufHdr(buf);
2166 :
2167 : /*
2168 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2169 : * worth storing the hashcode in BufferDesc so we need not recompute it
2170 : * here? Probably not.
2171 : */
2172 211334 : oldHash = BufTableHashCode(&oldTag);
2173 211334 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2174 :
2175 335382 : retry:
2176 :
2177 : /*
2178 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2179 : * association.
2180 : */
2181 335382 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2182 :
2183 : /* Re-lock the buffer header */
2184 335382 : buf_state = LockBufHdr(buf);
2185 :
2186 : /* If it's changed while we were waiting for lock, do nothing */
2187 335382 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2188 : {
2189 14 : UnlockBufHdr(buf);
2190 14 : LWLockRelease(oldPartitionLock);
2191 14 : return;
2192 : }
2193 :
2194 : /*
2195 : * We assume the reason for it to be pinned is that either we were
2196 : * asynchronously reading the page in before erroring out or someone else
2197 : * is flushing the page out. Wait for the IO to finish. (This could be
2198 : * an infinite loop if the refcount is messed up... it would be nice to
2199 : * time out after awhile, but there seems no way to be sure how many loops
2200 : * may be needed. Note that if the other guy has pinned the buffer but
2201 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2202 : * effectively be busy-looping here.)
2203 : */
2204 335368 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2205 : {
2206 124048 : UnlockBufHdr(buf);
2207 124048 : LWLockRelease(oldPartitionLock);
2208 : /* safety check: should definitely not be our *own* pin */
2209 124048 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2210 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2211 124048 : WaitIO(buf);
2212 124048 : goto retry;
2213 : }
2214 :
2215 : /*
2216 : * Clear out the buffer's tag and flags. We must do this to ensure that
2217 : * linear scans of the buffer array don't think the buffer is valid.
2218 : */
2219 211320 : oldFlags = buf_state & BUF_FLAG_MASK;
2220 211320 : ClearBufferTag(&buf->tag);
2221 :
2222 211320 : UnlockBufHdrExt(buf, buf_state,
2223 : 0,
2224 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2225 : 0);
2226 :
2227 : /*
2228 : * Remove the buffer from the lookup hashtable, if it was in there.
2229 : */
2230 211320 : if (oldFlags & BM_TAG_VALID)
2231 211320 : BufTableDelete(&oldTag, oldHash);
2232 :
2233 : /*
2234 : * Done with mapping lock.
2235 : */
2236 211320 : LWLockRelease(oldPartitionLock);
2237 : }
2238 :
2239 : /*
2240 : * Helper routine for GetVictimBuffer()
2241 : *
2242 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2243 : * buffer header spinlock held.
2244 : *
2245 : * Returns true if the buffer can be reused, in which case the buffer is only
2246 : * pinned by this backend and marked as invalid, false otherwise.
2247 : */
2248 : static bool
2249 2392124 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2250 : {
2251 : uint32 buf_state;
2252 : uint32 hash;
2253 : LWLock *partition_lock;
2254 : BufferTag tag;
2255 :
2256 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2257 :
2258 : /* have buffer pinned, so it's safe to read tag without lock */
2259 2392124 : tag = buf_hdr->tag;
2260 :
2261 2392124 : hash = BufTableHashCode(&tag);
2262 2392124 : partition_lock = BufMappingPartitionLock(hash);
2263 :
2264 2392124 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2265 :
2266 : /* lock the buffer header */
2267 2392124 : buf_state = LockBufHdr(buf_hdr);
2268 :
2269 : /*
2270 : * We have the buffer pinned nobody else should have been able to unset
2271 : * this concurrently.
2272 : */
2273 : Assert(buf_state & BM_TAG_VALID);
2274 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2275 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2276 :
2277 : /*
2278 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2279 : * give up on this buffer: It's clearly in use.
2280 : */
2281 2392124 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2282 : {
2283 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2284 :
2285 1040 : UnlockBufHdr(buf_hdr);
2286 1040 : LWLockRelease(partition_lock);
2287 :
2288 1040 : return false;
2289 : }
2290 :
2291 : /*
2292 : * Clear out the buffer's tag and flags and usagecount. This is not
2293 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2294 : * doing anything with the buffer. But currently it's beneficial, as the
2295 : * cheaper pre-check for several linear scans of shared buffers use the
2296 : * tag (see e.g. FlushDatabaseBuffers()).
2297 : */
2298 2391084 : ClearBufferTag(&buf_hdr->tag);
2299 2391084 : UnlockBufHdrExt(buf_hdr, buf_state,
2300 : 0,
2301 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2302 : 0);
2303 :
2304 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2305 :
2306 : /* finally delete buffer from the buffer mapping table */
2307 2391084 : BufTableDelete(&tag, hash);
2308 :
2309 2391084 : LWLockRelease(partition_lock);
2310 :
2311 2391084 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
2312 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2313 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2314 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u32(&buf_hdr->state)) > 0);
2315 :
2316 2391084 : return true;
2317 : }
2318 :
2319 : static Buffer
2320 3840556 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2321 : {
2322 : BufferDesc *buf_hdr;
2323 : Buffer buf;
2324 : uint32 buf_state;
2325 : bool from_ring;
2326 :
2327 : /*
2328 : * Ensure, before we pin a victim buffer, that there's a free refcount
2329 : * entry and resource owner slot for the pin.
2330 : */
2331 3840556 : ReservePrivateRefCountEntry();
2332 3840556 : ResourceOwnerEnlarge(CurrentResourceOwner);
2333 :
2334 : /* we return here if a prospective victim buffer gets used concurrently */
2335 13004 : again:
2336 :
2337 : /*
2338 : * Select a victim buffer. The buffer is returned pinned and owned by
2339 : * this backend.
2340 : */
2341 3853560 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2342 3853560 : buf = BufferDescriptorGetBuffer(buf_hdr);
2343 :
2344 : /*
2345 : * We shouldn't have any other pins for this buffer.
2346 : */
2347 3853560 : CheckBufferIsPinnedOnce(buf);
2348 :
2349 : /*
2350 : * If the buffer was dirty, try to write it out. There is a race
2351 : * condition here, in that someone might dirty it after we released the
2352 : * buffer header lock above, or even while we are writing it out (since
2353 : * our share-lock won't prevent hint-bit updates). We will recheck the
2354 : * dirty bit after re-locking the buffer header.
2355 : */
2356 3853560 : if (buf_state & BM_DIRTY)
2357 : {
2358 : LWLock *content_lock;
2359 :
2360 : Assert(buf_state & BM_TAG_VALID);
2361 : Assert(buf_state & BM_VALID);
2362 :
2363 : /*
2364 : * We need a share-lock on the buffer contents to write it out (else
2365 : * we might write invalid data, eg because someone else is compacting
2366 : * the page contents while we write). We must use a conditional lock
2367 : * acquisition here to avoid deadlock. Even though the buffer was not
2368 : * pinned (and therefore surely not locked) when StrategyGetBuffer
2369 : * returned it, someone else could have pinned and exclusive-locked it
2370 : * by the time we get here. If we try to get the lock unconditionally,
2371 : * we'd block waiting for them; if they later block waiting for us,
2372 : * deadlock ensues. (This has been observed to happen when two
2373 : * backends are both trying to split btree index pages, and the second
2374 : * one just happens to be trying to split the page the first one got
2375 : * from StrategyGetBuffer.)
2376 : */
2377 533792 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
2378 533792 : if (!LWLockConditionalAcquire(content_lock, LW_SHARED))
2379 : {
2380 : /*
2381 : * Someone else has locked the buffer, so give it up and loop back
2382 : * to get another one.
2383 : */
2384 0 : UnpinBuffer(buf_hdr);
2385 0 : goto again;
2386 : }
2387 :
2388 : /*
2389 : * If using a nondefault strategy, and writing the buffer would
2390 : * require a WAL flush, let the strategy decide whether to go ahead
2391 : * and write/reuse the buffer or to choose another victim. We need a
2392 : * lock to inspect the page LSN, so this can't be done inside
2393 : * StrategyGetBuffer.
2394 : */
2395 533792 : if (strategy != NULL)
2396 : {
2397 : XLogRecPtr lsn;
2398 :
2399 : /* Read the LSN while holding buffer header lock */
2400 156562 : buf_state = LockBufHdr(buf_hdr);
2401 156562 : lsn = BufferGetLSN(buf_hdr);
2402 156562 : UnlockBufHdr(buf_hdr);
2403 :
2404 156562 : if (XLogNeedsFlush(lsn)
2405 18380 : && StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2406 : {
2407 11964 : LWLockRelease(content_lock);
2408 11964 : UnpinBuffer(buf_hdr);
2409 11964 : goto again;
2410 : }
2411 : }
2412 :
2413 : /* OK, do the I/O */
2414 521828 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2415 521828 : LWLockRelease(content_lock);
2416 :
2417 521828 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2418 : &buf_hdr->tag);
2419 : }
2420 :
2421 :
2422 3841596 : if (buf_state & BM_VALID)
2423 : {
2424 : /*
2425 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2426 : * buffers are counted as IOOP_EVICT in the corresponding context
2427 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2428 : * strategy in two cases: 1) while initially claiming buffers for the
2429 : * strategy ring 2) to replace an existing strategy ring buffer
2430 : * because it is pinned or in use and cannot be reused.
2431 : *
2432 : * Blocks evicted from buffers already in the strategy ring are
2433 : * counted as IOOP_REUSE in the corresponding strategy context.
2434 : *
2435 : * At this point, we can accurately count evictions and reuses,
2436 : * because we have successfully claimed the valid buffer. Previously,
2437 : * we may have been forced to release the buffer due to concurrent
2438 : * pinners or erroring out.
2439 : */
2440 2387846 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2441 2387846 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2442 : }
2443 :
2444 : /*
2445 : * If the buffer has an entry in the buffer mapping table, delete it. This
2446 : * can fail because another backend could have pinned or dirtied the
2447 : * buffer.
2448 : */
2449 3841596 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2450 : {
2451 1040 : UnpinBuffer(buf_hdr);
2452 1040 : goto again;
2453 : }
2454 :
2455 : /* a final set of sanity checks */
2456 : #ifdef USE_ASSERT_CHECKING
2457 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
2458 :
2459 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2460 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2461 :
2462 : CheckBufferIsPinnedOnce(buf);
2463 : #endif
2464 :
2465 3840556 : return buf;
2466 : }
2467 :
2468 : /*
2469 : * Return the maximum number of buffers that a backend should try to pin once,
2470 : * to avoid exceeding its fair share. This is the highest value that
2471 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2472 : * system with a very small buffer pool relative to max_connections.
2473 : */
2474 : uint32
2475 1310634 : GetPinLimit(void)
2476 : {
2477 1310634 : return MaxProportionalPins;
2478 : }
2479 :
2480 : /*
2481 : * Return the maximum number of additional buffers that this backend should
2482 : * pin if it wants to stay under the per-backend limit, considering the number
2483 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2484 : * return by this function can be zero.
2485 : */
2486 : uint32
2487 7801766 : GetAdditionalPinLimit(void)
2488 : {
2489 : uint32 estimated_pins_held;
2490 :
2491 : /*
2492 : * We get the number of "overflowed" pins for free, but don't know the
2493 : * number of pins in PrivateRefCountArray. The cost of calculating that
2494 : * exactly doesn't seem worth it, so just assume the max.
2495 : */
2496 7801766 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2497 :
2498 : /* Is this backend already holding more than its fair share? */
2499 7801766 : if (estimated_pins_held > MaxProportionalPins)
2500 2385694 : return 0;
2501 :
2502 5416072 : return MaxProportionalPins - estimated_pins_held;
2503 : }
2504 :
2505 : /*
2506 : * Limit the number of pins a batch operation may additionally acquire, to
2507 : * avoid running out of pinnable buffers.
2508 : *
2509 : * One additional pin is always allowed, on the assumption that the operation
2510 : * requires at least one to make progress.
2511 : */
2512 : void
2513 404052 : LimitAdditionalPins(uint32 *additional_pins)
2514 : {
2515 : uint32 limit;
2516 :
2517 404052 : if (*additional_pins <= 1)
2518 384342 : return;
2519 :
2520 19710 : limit = GetAdditionalPinLimit();
2521 19710 : limit = Max(limit, 1);
2522 19710 : if (limit < *additional_pins)
2523 10766 : *additional_pins = limit;
2524 : }
2525 :
2526 : /*
2527 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2528 : * avoid duplicating the tracing and relpersistence related logic.
2529 : */
2530 : static BlockNumber
2531 426870 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2532 : ForkNumber fork,
2533 : BufferAccessStrategy strategy,
2534 : uint32 flags,
2535 : uint32 extend_by,
2536 : BlockNumber extend_upto,
2537 : Buffer *buffers,
2538 : uint32 *extended_by)
2539 : {
2540 : BlockNumber first_block;
2541 :
2542 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2543 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2544 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2545 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2546 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2547 : extend_by);
2548 :
2549 426870 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2550 22818 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2551 : extend_by, extend_upto,
2552 : buffers, &extend_by);
2553 : else
2554 404052 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2555 : extend_by, extend_upto,
2556 : buffers, &extend_by);
2557 426870 : *extended_by = extend_by;
2558 :
2559 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2560 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2561 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2562 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2563 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2564 : *extended_by,
2565 : first_block);
2566 :
2567 426870 : return first_block;
2568 : }
2569 :
2570 : /*
2571 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2572 : * shared buffers.
2573 : */
2574 : static BlockNumber
2575 404052 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2576 : ForkNumber fork,
2577 : BufferAccessStrategy strategy,
2578 : uint32 flags,
2579 : uint32 extend_by,
2580 : BlockNumber extend_upto,
2581 : Buffer *buffers,
2582 : uint32 *extended_by)
2583 : {
2584 : BlockNumber first_block;
2585 404052 : IOContext io_context = IOContextForStrategy(strategy);
2586 : instr_time io_start;
2587 :
2588 404052 : LimitAdditionalPins(&extend_by);
2589 :
2590 : /*
2591 : * Acquire victim buffers for extension without holding extension lock.
2592 : * Writing out victim buffers is the most expensive part of extending the
2593 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2594 : * the buffers is also quite expensive, so do that before holding the
2595 : * extension lock as well.
2596 : *
2597 : * These pages are pinned by us and not valid. While we hold the pin they
2598 : * can't be acquired as victim buffers by another backend.
2599 : */
2600 850494 : for (uint32 i = 0; i < extend_by; i++)
2601 : {
2602 : Block buf_block;
2603 :
2604 446442 : buffers[i] = GetVictimBuffer(strategy, io_context);
2605 446442 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2606 :
2607 : /* new buffers are zero-filled */
2608 446442 : MemSet(buf_block, 0, BLCKSZ);
2609 : }
2610 :
2611 : /*
2612 : * Lock relation against concurrent extensions, unless requested not to.
2613 : *
2614 : * We use the same extension lock for all forks. That's unnecessarily
2615 : * restrictive, but currently extensions for forks don't happen often
2616 : * enough to make it worth locking more granularly.
2617 : *
2618 : * Note that another backend might have extended the relation by the time
2619 : * we get the lock.
2620 : */
2621 404052 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2622 300980 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2623 :
2624 : /*
2625 : * If requested, invalidate size cache, so that smgrnblocks asks the
2626 : * kernel.
2627 : */
2628 404052 : if (flags & EB_CLEAR_SIZE_CACHE)
2629 15468 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2630 :
2631 404052 : first_block = smgrnblocks(BMR_GET_SMGR(bmr), fork);
2632 :
2633 : /*
2634 : * Now that we have the accurate relation size, check if the caller wants
2635 : * us to extend to only up to a specific size. If there were concurrent
2636 : * extensions, we might have acquired too many buffers and need to release
2637 : * them.
2638 : */
2639 404052 : if (extend_upto != InvalidBlockNumber)
2640 : {
2641 106688 : uint32 orig_extend_by = extend_by;
2642 :
2643 106688 : if (first_block > extend_upto)
2644 0 : extend_by = 0;
2645 106688 : else if ((uint64) first_block + extend_by > extend_upto)
2646 20 : extend_by = extend_upto - first_block;
2647 :
2648 106730 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2649 : {
2650 42 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2651 :
2652 42 : UnpinBuffer(buf_hdr);
2653 : }
2654 :
2655 106688 : if (extend_by == 0)
2656 : {
2657 20 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2658 20 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2659 20 : *extended_by = extend_by;
2660 20 : return first_block;
2661 : }
2662 : }
2663 :
2664 : /* Fail if relation is already at maximum possible length */
2665 404032 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2666 0 : ereport(ERROR,
2667 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2668 : errmsg("cannot extend relation %s beyond %u blocks",
2669 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str,
2670 : MaxBlockNumber)));
2671 :
2672 : /*
2673 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2674 : *
2675 : * This needs to happen before we extend the relation, because as soon as
2676 : * we do, other backends can start to read in those pages.
2677 : */
2678 850432 : for (uint32 i = 0; i < extend_by; i++)
2679 : {
2680 446400 : Buffer victim_buf = buffers[i];
2681 446400 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2682 : BufferTag tag;
2683 : uint32 hash;
2684 : LWLock *partition_lock;
2685 : int existing_id;
2686 :
2687 : /* in case we need to pin an existing buffer below */
2688 446400 : ResourceOwnerEnlarge(CurrentResourceOwner);
2689 446400 : ReservePrivateRefCountEntry();
2690 :
2691 446400 : InitBufferTag(&tag, &BMR_GET_SMGR(bmr)->smgr_rlocator.locator, fork,
2692 : first_block + i);
2693 446400 : hash = BufTableHashCode(&tag);
2694 446400 : partition_lock = BufMappingPartitionLock(hash);
2695 :
2696 446400 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2697 :
2698 446400 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2699 :
2700 : /*
2701 : * We get here only in the corner case where we are trying to extend
2702 : * the relation but we found a pre-existing buffer. This can happen
2703 : * because a prior attempt at extending the relation failed, and
2704 : * because mdread doesn't complain about reads beyond EOF (when
2705 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2706 : * beyond EOF could have left a "valid" zero-filled buffer.
2707 : *
2708 : * This has also been observed when relation was overwritten by
2709 : * external process. Since the legitimate cases should always have
2710 : * left a zero-filled buffer, complain if not PageIsNew.
2711 : */
2712 446400 : if (existing_id >= 0)
2713 : {
2714 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2715 : Block buf_block;
2716 : bool valid;
2717 :
2718 : /*
2719 : * Pin the existing buffer before releasing the partition lock,
2720 : * preventing it from being evicted.
2721 : */
2722 0 : valid = PinBuffer(existing_hdr, strategy, false);
2723 :
2724 0 : LWLockRelease(partition_lock);
2725 0 : UnpinBuffer(victim_buf_hdr);
2726 :
2727 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2728 0 : buf_block = BufHdrGetBlock(existing_hdr);
2729 :
2730 0 : if (valid && !PageIsNew((Page) buf_block))
2731 0 : ereport(ERROR,
2732 : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2733 : existing_hdr->tag.blockNum,
2734 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str)));
2735 :
2736 : /*
2737 : * We *must* do smgr[zero]extend before succeeding, else the page
2738 : * will not be reserved by the kernel, and the next P_NEW call
2739 : * will decide to return the same page. Clear the BM_VALID bit,
2740 : * do StartBufferIO() and proceed.
2741 : *
2742 : * Loop to handle the very small possibility that someone re-sets
2743 : * BM_VALID between our clearing it and StartBufferIO inspecting
2744 : * it.
2745 : */
2746 : do
2747 : {
2748 0 : pg_atomic_fetch_and_u32(&existing_hdr->state, ~BM_VALID);
2749 0 : } while (!StartBufferIO(existing_hdr, true, false));
2750 : }
2751 : else
2752 : {
2753 : uint32 buf_state;
2754 446400 : uint32 set_bits = 0;
2755 :
2756 446400 : buf_state = LockBufHdr(victim_buf_hdr);
2757 :
2758 : /* some sanity checks while we hold the buffer header lock */
2759 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY | BM_JUST_DIRTIED)));
2760 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2761 :
2762 446400 : victim_buf_hdr->tag = tag;
2763 :
2764 446400 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2765 446400 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2766 435794 : set_bits |= BM_PERMANENT;
2767 :
2768 446400 : UnlockBufHdrExt(victim_buf_hdr, buf_state,
2769 : set_bits, 0,
2770 : 0);
2771 :
2772 446400 : LWLockRelease(partition_lock);
2773 :
2774 : /* XXX: could combine the locked operations in it with the above */
2775 446400 : StartBufferIO(victim_buf_hdr, true, false);
2776 : }
2777 : }
2778 :
2779 404032 : io_start = pgstat_prepare_io_time(track_io_timing);
2780 :
2781 : /*
2782 : * Note: if smgrzeroextend fails, we will end up with buffers that are
2783 : * allocated but not marked BM_VALID. The next relation extension will
2784 : * still select the same block number (because the relation didn't get any
2785 : * longer on disk) and so future attempts to extend the relation will find
2786 : * the same buffers (if they have not been recycled) but come right back
2787 : * here to try smgrzeroextend again.
2788 : *
2789 : * We don't need to set checksum for all-zero pages.
2790 : */
2791 404032 : smgrzeroextend(BMR_GET_SMGR(bmr), fork, first_block, extend_by, false);
2792 :
2793 : /*
2794 : * Release the file-extension lock; it's now OK for someone else to extend
2795 : * the relation some more.
2796 : *
2797 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
2798 : * take noticeable time.
2799 : */
2800 404032 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2801 300960 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2802 :
2803 404032 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
2804 404032 : io_start, 1, extend_by * BLCKSZ);
2805 :
2806 : /* Set BM_VALID, terminate IO, and wake up any waiters */
2807 850432 : for (uint32 i = 0; i < extend_by; i++)
2808 : {
2809 446400 : Buffer buf = buffers[i];
2810 446400 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
2811 446400 : bool lock = false;
2812 :
2813 446400 : if (flags & EB_LOCK_FIRST && i == 0)
2814 296830 : lock = true;
2815 149570 : else if (flags & EB_LOCK_TARGET)
2816 : {
2817 : Assert(extend_upto != InvalidBlockNumber);
2818 89086 : if (first_block + i + 1 == extend_upto)
2819 88158 : lock = true;
2820 : }
2821 :
2822 446400 : if (lock)
2823 384988 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2824 :
2825 446400 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
2826 : }
2827 :
2828 404032 : pgBufferUsage.shared_blks_written += extend_by;
2829 :
2830 404032 : *extended_by = extend_by;
2831 :
2832 404032 : return first_block;
2833 : }
2834 :
2835 : /*
2836 : * BufferIsLockedByMe
2837 : *
2838 : * Checks if this backend has the buffer locked in any mode.
2839 : *
2840 : * Buffer must be pinned.
2841 : */
2842 : bool
2843 0 : BufferIsLockedByMe(Buffer buffer)
2844 : {
2845 : BufferDesc *bufHdr;
2846 :
2847 : Assert(BufferIsPinned(buffer));
2848 :
2849 0 : if (BufferIsLocal(buffer))
2850 : {
2851 : /* Content locks are not maintained for local buffers. */
2852 0 : return true;
2853 : }
2854 : else
2855 : {
2856 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2857 0 : return LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr));
2858 : }
2859 : }
2860 :
2861 : /*
2862 : * BufferIsLockedByMeInMode
2863 : *
2864 : * Checks if this backend has the buffer locked in the specified mode.
2865 : *
2866 : * Buffer must be pinned.
2867 : */
2868 : bool
2869 0 : BufferIsLockedByMeInMode(Buffer buffer, int mode)
2870 : {
2871 : BufferDesc *bufHdr;
2872 :
2873 : Assert(BufferIsPinned(buffer));
2874 :
2875 0 : if (BufferIsLocal(buffer))
2876 : {
2877 : /* Content locks are not maintained for local buffers. */
2878 0 : return true;
2879 : }
2880 : else
2881 : {
2882 : LWLockMode lw_mode;
2883 :
2884 0 : switch (mode)
2885 : {
2886 0 : case BUFFER_LOCK_EXCLUSIVE:
2887 0 : lw_mode = LW_EXCLUSIVE;
2888 0 : break;
2889 0 : case BUFFER_LOCK_SHARE:
2890 0 : lw_mode = LW_SHARED;
2891 0 : break;
2892 0 : default:
2893 0 : pg_unreachable();
2894 : }
2895 :
2896 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2897 0 : return LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2898 : lw_mode);
2899 : }
2900 : }
2901 :
2902 : /*
2903 : * BufferIsDirty
2904 : *
2905 : * Checks if buffer is already dirty.
2906 : *
2907 : * Buffer must be pinned and exclusive-locked. (Without an exclusive lock,
2908 : * the result may be stale before it's returned.)
2909 : */
2910 : bool
2911 0 : BufferIsDirty(Buffer buffer)
2912 : {
2913 : BufferDesc *bufHdr;
2914 :
2915 : Assert(BufferIsPinned(buffer));
2916 :
2917 0 : if (BufferIsLocal(buffer))
2918 : {
2919 0 : int bufid = -buffer - 1;
2920 :
2921 0 : bufHdr = GetLocalBufferDescriptor(bufid);
2922 : /* Content locks are not maintained for local buffers. */
2923 : }
2924 : else
2925 : {
2926 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2927 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
2928 : }
2929 :
2930 0 : return pg_atomic_read_u32(&bufHdr->state) & BM_DIRTY;
2931 : }
2932 :
2933 : /*
2934 : * MarkBufferDirty
2935 : *
2936 : * Marks buffer contents as dirty (actual write happens later).
2937 : *
2938 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
2939 : * exclusive lock, then somebody could be in process of writing the buffer,
2940 : * leading to risk of bad data written to disk.)
2941 : */
2942 : void
2943 43044904 : MarkBufferDirty(Buffer buffer)
2944 : {
2945 : BufferDesc *bufHdr;
2946 : uint32 buf_state;
2947 : uint32 old_buf_state;
2948 :
2949 43044904 : if (!BufferIsValid(buffer))
2950 0 : elog(ERROR, "bad buffer ID: %d", buffer);
2951 :
2952 43044904 : if (BufferIsLocal(buffer))
2953 : {
2954 2444466 : MarkLocalBufferDirty(buffer);
2955 2444466 : return;
2956 : }
2957 :
2958 40600438 : bufHdr = GetBufferDescriptor(buffer - 1);
2959 :
2960 : Assert(BufferIsPinned(buffer));
2961 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
2962 :
2963 : /*
2964 : * NB: We have to wait for the buffer header spinlock to be not held, as
2965 : * TerminateBufferIO() relies on the spinlock.
2966 : */
2967 40600438 : old_buf_state = pg_atomic_read_u32(&bufHdr->state);
2968 : for (;;)
2969 : {
2970 40600624 : if (old_buf_state & BM_LOCKED)
2971 2 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
2972 :
2973 40600624 : buf_state = old_buf_state;
2974 :
2975 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2976 40600624 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
2977 :
2978 40600624 : if (pg_atomic_compare_exchange_u32(&bufHdr->state, &old_buf_state,
2979 : buf_state))
2980 40600438 : break;
2981 : }
2982 :
2983 : /*
2984 : * If the buffer was not dirty already, do vacuum accounting.
2985 : */
2986 40600438 : if (!(old_buf_state & BM_DIRTY))
2987 : {
2988 1301366 : pgBufferUsage.shared_blks_dirtied++;
2989 1301366 : if (VacuumCostActive)
2990 13542 : VacuumCostBalance += VacuumCostPageDirty;
2991 : }
2992 : }
2993 :
2994 : /*
2995 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
2996 : *
2997 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
2998 : * compared to calling the two routines separately. Now it's mainly just
2999 : * a convenience function. However, if the passed buffer is valid and
3000 : * already contains the desired block, we just return it as-is; and that
3001 : * does save considerable work compared to a full release and reacquire.
3002 : *
3003 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
3004 : * buffer actually needs to be released. This case is the same as ReadBuffer,
3005 : * but can save some tests in the caller.
3006 : */
3007 : Buffer
3008 57960734 : ReleaseAndReadBuffer(Buffer buffer,
3009 : Relation relation,
3010 : BlockNumber blockNum)
3011 : {
3012 57960734 : ForkNumber forkNum = MAIN_FORKNUM;
3013 : BufferDesc *bufHdr;
3014 :
3015 57960734 : if (BufferIsValid(buffer))
3016 : {
3017 : Assert(BufferIsPinned(buffer));
3018 34642846 : if (BufferIsLocal(buffer))
3019 : {
3020 73728 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3021 80772 : if (bufHdr->tag.blockNum == blockNum &&
3022 14088 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3023 7044 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3024 7044 : return buffer;
3025 66684 : UnpinLocalBuffer(buffer);
3026 : }
3027 : else
3028 : {
3029 34569118 : bufHdr = GetBufferDescriptor(buffer - 1);
3030 : /* we have pin, so it's ok to examine tag without spinlock */
3031 46414456 : if (bufHdr->tag.blockNum == blockNum &&
3032 23690676 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3033 11845338 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3034 11845338 : return buffer;
3035 22723780 : UnpinBuffer(bufHdr);
3036 : }
3037 : }
3038 :
3039 46108352 : return ReadBuffer(relation, blockNum);
3040 : }
3041 :
3042 : /*
3043 : * PinBuffer -- make buffer unavailable for replacement.
3044 : *
3045 : * For the default access strategy, the buffer's usage_count is incremented
3046 : * when we first pin it; for other strategies we just make sure the usage_count
3047 : * isn't zero. (The idea of the latter is that we don't want synchronized
3048 : * heap scans to inflate the count, but we need it to not be zero to discourage
3049 : * other backends from stealing buffers from our ring. As long as we cycle
3050 : * through the ring faster than the global clock-sweep cycles, buffers in
3051 : * our ring won't be chosen as victims for replacement by other backends.)
3052 : *
3053 : * This should be applied only to shared buffers, never local ones.
3054 : *
3055 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3056 : * taking the buffer header lock; instead update the state variable in loop of
3057 : * CAS operations. Hopefully it's just a single CAS.
3058 : *
3059 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3060 : * must have been done already.
3061 : *
3062 : * Returns true if buffer is BM_VALID, else false. This provision allows
3063 : * some callers to avoid an extra spinlock cycle. If skip_if_not_valid is
3064 : * true, then a false return value also indicates that the buffer was
3065 : * (recently) invalid and has not been pinned.
3066 : */
3067 : static bool
3068 118456732 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
3069 : bool skip_if_not_valid)
3070 : {
3071 118456732 : Buffer b = BufferDescriptorGetBuffer(buf);
3072 : bool result;
3073 : PrivateRefCountEntry *ref;
3074 :
3075 : Assert(!BufferIsLocal(b));
3076 : Assert(ReservedRefCountEntry != NULL);
3077 :
3078 118456732 : ref = GetPrivateRefCountEntry(b, true);
3079 :
3080 118456732 : if (ref == NULL)
3081 : {
3082 : uint32 buf_state;
3083 : uint32 old_buf_state;
3084 :
3085 114125298 : old_buf_state = pg_atomic_read_u32(&buf->state);
3086 : for (;;)
3087 : {
3088 114166910 : if (unlikely(skip_if_not_valid && !(old_buf_state & BM_VALID)))
3089 12 : return false;
3090 :
3091 : /*
3092 : * We're not allowed to increase the refcount while the buffer
3093 : * header spinlock is held. Wait for the lock to be released.
3094 : */
3095 114166898 : if (old_buf_state & BM_LOCKED)
3096 1682 : old_buf_state = WaitBufHdrUnlocked(buf);
3097 :
3098 114166898 : buf_state = old_buf_state;
3099 :
3100 : /* increase refcount */
3101 114166898 : buf_state += BUF_REFCOUNT_ONE;
3102 :
3103 114166898 : if (strategy == NULL)
3104 : {
3105 : /* Default case: increase usagecount unless already max. */
3106 112611950 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3107 6611126 : buf_state += BUF_USAGECOUNT_ONE;
3108 : }
3109 : else
3110 : {
3111 : /*
3112 : * Ring buffers shouldn't evict others from pool. Thus we
3113 : * don't make usagecount more than 1.
3114 : */
3115 1554948 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3116 62208 : buf_state += BUF_USAGECOUNT_ONE;
3117 : }
3118 :
3119 114166898 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3120 : buf_state))
3121 : {
3122 114125286 : result = (buf_state & BM_VALID) != 0;
3123 :
3124 114125286 : TrackNewBufferPin(b);
3125 114125286 : break;
3126 : }
3127 : }
3128 : }
3129 : else
3130 : {
3131 : /*
3132 : * If we previously pinned the buffer, it is likely to be valid, but
3133 : * it may not be if StartReadBuffers() was called and
3134 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3135 : * the flags without locking. This is racy, but it's OK to return
3136 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3137 : * it'll see that it's now valid.
3138 : *
3139 : * Note: We deliberately avoid a Valgrind client request here.
3140 : * Individual access methods can optionally superimpose buffer page
3141 : * client requests on top of our client requests to enforce that
3142 : * buffers are only accessed while locked (and pinned). It's possible
3143 : * that the buffer page is legitimately non-accessible here. We
3144 : * cannot meddle with that.
3145 : */
3146 4331434 : result = (pg_atomic_read_u32(&buf->state) & BM_VALID) != 0;
3147 :
3148 : Assert(ref->refcount > 0);
3149 4331434 : ref->refcount++;
3150 4331434 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3151 : }
3152 :
3153 118456720 : return result;
3154 : }
3155 :
3156 : /*
3157 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3158 : * The spinlock is released before return.
3159 : *
3160 : * As this function is called with the spinlock held, the caller has to
3161 : * previously call ReservePrivateRefCountEntry() and
3162 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3163 : *
3164 : * Currently, no callers of this function want to modify the buffer's
3165 : * usage_count at all, so there's no need for a strategy parameter.
3166 : * Also we don't bother with a BM_VALID test (the caller could check that for
3167 : * itself).
3168 : *
3169 : * Also all callers only ever use this function when it's known that the
3170 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3171 : * searching the private refcount array & hash, which is a boon, because the
3172 : * spinlock is still held.
3173 : *
3174 : * Note: use of this routine is frequently mandatory, not just an optimization
3175 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3176 : * its state can change under us.
3177 : */
3178 : static void
3179 603080 : PinBuffer_Locked(BufferDesc *buf)
3180 : {
3181 : uint32 old_buf_state;
3182 :
3183 : /*
3184 : * As explained, We don't expect any preexisting pins. That allows us to
3185 : * manipulate the PrivateRefCount after releasing the spinlock
3186 : */
3187 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3188 :
3189 : /*
3190 : * Since we hold the buffer spinlock, we can update the buffer state and
3191 : * release the lock in one operation.
3192 : */
3193 603080 : old_buf_state = pg_atomic_read_u32(&buf->state);
3194 :
3195 603080 : UnlockBufHdrExt(buf, old_buf_state,
3196 : 0, 0, 1);
3197 :
3198 603080 : TrackNewBufferPin(BufferDescriptorGetBuffer(buf));
3199 603080 : }
3200 :
3201 : /*
3202 : * Support for waking up another backend that is waiting for the cleanup lock
3203 : * to be released using BM_PIN_COUNT_WAITER.
3204 : *
3205 : * See LockBufferForCleanup().
3206 : *
3207 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3208 : * not just reducing the backend-local pincount for the buffer).
3209 : */
3210 : static void
3211 24 : WakePinCountWaiter(BufferDesc *buf)
3212 : {
3213 : /*
3214 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3215 : * backend could have unpinned this buffer, and already woken up the
3216 : * waiter.
3217 : *
3218 : * There's no danger of the buffer being replaced after we unpinned it
3219 : * above, as it's pinned by the waiter. The waiter removes
3220 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3221 : * backend waking it up.
3222 : */
3223 24 : uint32 buf_state = LockBufHdr(buf);
3224 :
3225 24 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3226 24 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3227 24 : {
3228 : /* we just released the last pin other than the waiter's */
3229 24 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3230 :
3231 24 : UnlockBufHdrExt(buf, buf_state,
3232 : 0, BM_PIN_COUNT_WAITER,
3233 : 0);
3234 24 : ProcSendSignal(wait_backend_pgprocno);
3235 : }
3236 : else
3237 0 : UnlockBufHdr(buf);
3238 24 : }
3239 :
3240 : /*
3241 : * UnpinBuffer -- make buffer available for replacement.
3242 : *
3243 : * This should be applied only to shared buffers, never local ones. This
3244 : * always adjusts CurrentResourceOwner.
3245 : */
3246 : static void
3247 145564836 : UnpinBuffer(BufferDesc *buf)
3248 : {
3249 145564836 : Buffer b = BufferDescriptorGetBuffer(buf);
3250 :
3251 145564836 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3252 145564836 : UnpinBufferNoOwner(buf);
3253 145564836 : }
3254 :
3255 : static void
3256 145573932 : UnpinBufferNoOwner(BufferDesc *buf)
3257 : {
3258 : PrivateRefCountEntry *ref;
3259 145573932 : Buffer b = BufferDescriptorGetBuffer(buf);
3260 :
3261 : Assert(!BufferIsLocal(b));
3262 :
3263 : /* not moving as we're likely deleting it soon anyway */
3264 145573932 : ref = GetPrivateRefCountEntry(b, false);
3265 : Assert(ref != NULL);
3266 : Assert(ref->refcount > 0);
3267 145573932 : ref->refcount--;
3268 145573932 : if (ref->refcount == 0)
3269 : {
3270 : uint32 old_buf_state;
3271 :
3272 : /*
3273 : * Mark buffer non-accessible to Valgrind.
3274 : *
3275 : * Note that the buffer may have already been marked non-accessible
3276 : * within access method code that enforces that buffers are only
3277 : * accessed while a buffer lock is held.
3278 : */
3279 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3280 :
3281 : /*
3282 : * I'd better not still hold the buffer content lock. Can't use
3283 : * BufferIsLockedByMe(), as that asserts the buffer is pinned.
3284 : */
3285 : Assert(!LWLockHeldByMe(BufferDescriptorGetContentLock(buf)));
3286 :
3287 : /* decrement the shared reference count */
3288 118581926 : old_buf_state = pg_atomic_fetch_sub_u32(&buf->state, BUF_REFCOUNT_ONE);
3289 :
3290 : /* Support LockBufferForCleanup() */
3291 118581926 : if (old_buf_state & BM_PIN_COUNT_WAITER)
3292 24 : WakePinCountWaiter(buf);
3293 :
3294 118581926 : ForgetPrivateRefCountEntry(ref);
3295 : }
3296 145573932 : }
3297 :
3298 : /*
3299 : * Set up backend-local tracking of a buffer pinned the first time by this
3300 : * backend.
3301 : */
3302 : inline void
3303 118581926 : TrackNewBufferPin(Buffer buf)
3304 : {
3305 : PrivateRefCountEntry *ref;
3306 :
3307 118581926 : ref = NewPrivateRefCountEntry(buf);
3308 118581926 : ref->refcount++;
3309 :
3310 118581926 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buf);
3311 :
3312 : /*
3313 : * This is the first pin for this page by this backend, mark its page as
3314 : * defined to valgrind. While the page contents might not actually be
3315 : * valid yet, we don't currently guarantee that such pages are marked
3316 : * undefined or non-accessible.
3317 : *
3318 : * It's not necessarily the prettiest to do this here, but otherwise we'd
3319 : * need this block of code in multiple places.
3320 : */
3321 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(GetBufferDescriptor(buf - 1)),
3322 : BLCKSZ);
3323 118581926 : }
3324 :
3325 : #define ST_SORT sort_checkpoint_bufferids
3326 : #define ST_ELEMENT_TYPE CkptSortItem
3327 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3328 : #define ST_SCOPE static
3329 : #define ST_DEFINE
3330 : #include "lib/sort_template.h"
3331 :
3332 : /*
3333 : * BufferSync -- Write out all dirty buffers in the pool.
3334 : *
3335 : * This is called at checkpoint time to write out all dirty shared buffers.
3336 : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3337 : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3338 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3339 : * even unlogged buffers, which are otherwise skipped. The remaining flags
3340 : * currently have no effect here.
3341 : */
3342 : static void
3343 3466 : BufferSync(int flags)
3344 : {
3345 : uint32 buf_state;
3346 : int buf_id;
3347 : int num_to_scan;
3348 : int num_spaces;
3349 : int num_processed;
3350 : int num_written;
3351 3466 : CkptTsStatus *per_ts_stat = NULL;
3352 : Oid last_tsid;
3353 : binaryheap *ts_heap;
3354 : int i;
3355 3466 : uint32 mask = BM_DIRTY;
3356 : WritebackContext wb_context;
3357 :
3358 : /*
3359 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3360 : * we write only permanent, dirty buffers. But at shutdown or end of
3361 : * recovery, we write all dirty buffers.
3362 : */
3363 3466 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3364 : CHECKPOINT_FLUSH_UNLOGGED))))
3365 1946 : mask |= BM_PERMANENT;
3366 :
3367 : /*
3368 : * Loop over all buffers, and mark the ones that need to be written with
3369 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3370 : * can estimate how much work needs to be done.
3371 : *
3372 : * This allows us to write only those pages that were dirty when the
3373 : * checkpoint began, and not those that get dirtied while it proceeds.
3374 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3375 : * later in this function, or by normal backends or the bgwriter cleaning
3376 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3377 : * have the flag set.
3378 : *
3379 : * Note that if we fail to write some buffer, we may leave buffers with
3380 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3381 : * certainly need to be written for the next checkpoint attempt, too.
3382 : */
3383 3466 : num_to_scan = 0;
3384 24115242 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3385 : {
3386 24111776 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3387 24111776 : uint32 set_bits = 0;
3388 :
3389 : /*
3390 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3391 : * SyncOneBuffer.
3392 : */
3393 24111776 : buf_state = LockBufHdr(bufHdr);
3394 :
3395 24111776 : if ((buf_state & mask) == mask)
3396 : {
3397 : CkptSortItem *item;
3398 :
3399 584386 : set_bits = BM_CHECKPOINT_NEEDED;
3400 :
3401 584386 : item = &CkptBufferIds[num_to_scan++];
3402 584386 : item->buf_id = buf_id;
3403 584386 : item->tsId = bufHdr->tag.spcOid;
3404 584386 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3405 584386 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3406 584386 : item->blockNum = bufHdr->tag.blockNum;
3407 : }
3408 :
3409 24111776 : UnlockBufHdrExt(bufHdr, buf_state,
3410 : set_bits, 0,
3411 : 0);
3412 :
3413 : /* Check for barrier events in case NBuffers is large. */
3414 24111776 : if (ProcSignalBarrierPending)
3415 0 : ProcessProcSignalBarrier();
3416 : }
3417 :
3418 3466 : if (num_to_scan == 0)
3419 1328 : return; /* nothing to do */
3420 :
3421 2138 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3422 :
3423 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3424 :
3425 : /*
3426 : * Sort buffers that need to be written to reduce the likelihood of random
3427 : * IO. The sorting is also important for the implementation of balancing
3428 : * writes between tablespaces. Without balancing writes we'd potentially
3429 : * end up writing to the tablespaces one-by-one; possibly overloading the
3430 : * underlying system.
3431 : */
3432 2138 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3433 :
3434 2138 : num_spaces = 0;
3435 :
3436 : /*
3437 : * Allocate progress status for each tablespace with buffers that need to
3438 : * be flushed. This requires the to-be-flushed array to be sorted.
3439 : */
3440 2138 : last_tsid = InvalidOid;
3441 586524 : for (i = 0; i < num_to_scan; i++)
3442 : {
3443 : CkptTsStatus *s;
3444 : Oid cur_tsid;
3445 :
3446 584386 : cur_tsid = CkptBufferIds[i].tsId;
3447 :
3448 : /*
3449 : * Grow array of per-tablespace status structs, every time a new
3450 : * tablespace is found.
3451 : */
3452 584386 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3453 3196 : {
3454 : Size sz;
3455 :
3456 3196 : num_spaces++;
3457 :
3458 : /*
3459 : * Not worth adding grow-by-power-of-2 logic here - even with a
3460 : * few hundred tablespaces this should be fine.
3461 : */
3462 3196 : sz = sizeof(CkptTsStatus) * num_spaces;
3463 :
3464 3196 : if (per_ts_stat == NULL)
3465 2138 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3466 : else
3467 1058 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3468 :
3469 3196 : s = &per_ts_stat[num_spaces - 1];
3470 3196 : memset(s, 0, sizeof(*s));
3471 3196 : s->tsId = cur_tsid;
3472 :
3473 : /*
3474 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3475 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3476 : * will follow afterwards.
3477 : */
3478 3196 : s->index = i;
3479 :
3480 : /*
3481 : * progress_slice will be determined once we know how many buffers
3482 : * are in each tablespace, i.e. after this loop.
3483 : */
3484 :
3485 3196 : last_tsid = cur_tsid;
3486 : }
3487 : else
3488 : {
3489 581190 : s = &per_ts_stat[num_spaces - 1];
3490 : }
3491 :
3492 584386 : s->num_to_scan++;
3493 :
3494 : /* Check for barrier events. */
3495 584386 : if (ProcSignalBarrierPending)
3496 0 : ProcessProcSignalBarrier();
3497 : }
3498 :
3499 : Assert(num_spaces > 0);
3500 :
3501 : /*
3502 : * Build a min-heap over the write-progress in the individual tablespaces,
3503 : * and compute how large a portion of the total progress a single
3504 : * processed buffer is.
3505 : */
3506 2138 : ts_heap = binaryheap_allocate(num_spaces,
3507 : ts_ckpt_progress_comparator,
3508 : NULL);
3509 :
3510 5334 : for (i = 0; i < num_spaces; i++)
3511 : {
3512 3196 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3513 :
3514 3196 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3515 :
3516 3196 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3517 : }
3518 :
3519 2138 : binaryheap_build(ts_heap);
3520 :
3521 : /*
3522 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3523 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3524 : * tablespaces; otherwise the sorting would lead to only one tablespace
3525 : * receiving writes at a time, making inefficient use of the hardware.
3526 : */
3527 2138 : num_processed = 0;
3528 2138 : num_written = 0;
3529 586524 : while (!binaryheap_empty(ts_heap))
3530 : {
3531 584386 : BufferDesc *bufHdr = NULL;
3532 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3533 584386 : DatumGetPointer(binaryheap_first(ts_heap));
3534 :
3535 584386 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3536 : Assert(buf_id != -1);
3537 :
3538 584386 : bufHdr = GetBufferDescriptor(buf_id);
3539 :
3540 584386 : num_processed++;
3541 :
3542 : /*
3543 : * We don't need to acquire the lock here, because we're only looking
3544 : * at a single bit. It's possible that someone else writes the buffer
3545 : * and clears the flag right after we check, but that doesn't matter
3546 : * since SyncOneBuffer will then do nothing. However, there is a
3547 : * further race condition: it's conceivable that between the time we
3548 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3549 : * someone else not only wrote the buffer but replaced it with another
3550 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3551 : * write the buffer though we didn't need to. It doesn't seem worth
3552 : * guarding against this, though.
3553 : */
3554 584386 : if (pg_atomic_read_u32(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3555 : {
3556 544928 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3557 : {
3558 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3559 544928 : PendingCheckpointerStats.buffers_written++;
3560 544928 : num_written++;
3561 : }
3562 : }
3563 :
3564 : /*
3565 : * Measure progress independent of actually having to flush the buffer
3566 : * - otherwise writing become unbalanced.
3567 : */
3568 584386 : ts_stat->progress += ts_stat->progress_slice;
3569 584386 : ts_stat->num_scanned++;
3570 584386 : ts_stat->index++;
3571 :
3572 : /* Have all the buffers from the tablespace been processed? */
3573 584386 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3574 : {
3575 3196 : binaryheap_remove_first(ts_heap);
3576 : }
3577 : else
3578 : {
3579 : /* update heap with the new progress */
3580 581190 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3581 : }
3582 :
3583 : /*
3584 : * Sleep to throttle our I/O rate.
3585 : *
3586 : * (This will check for barrier events even if it doesn't sleep.)
3587 : */
3588 584386 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3589 : }
3590 :
3591 : /*
3592 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3593 : * IOContext will always be IOCONTEXT_NORMAL.
3594 : */
3595 2138 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3596 :
3597 2138 : pfree(per_ts_stat);
3598 2138 : per_ts_stat = NULL;
3599 2138 : binaryheap_free(ts_heap);
3600 :
3601 : /*
3602 : * Update checkpoint statistics. As noted above, this doesn't include
3603 : * buffers written by other backends or bgwriter scan.
3604 : */
3605 2138 : CheckpointStats.ckpt_bufs_written += num_written;
3606 :
3607 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3608 : }
3609 :
3610 : /*
3611 : * BgBufferSync -- Write out some dirty buffers in the pool.
3612 : *
3613 : * This is called periodically by the background writer process.
3614 : *
3615 : * Returns true if it's appropriate for the bgwriter process to go into
3616 : * low-power hibernation mode. (This happens if the strategy clock-sweep
3617 : * has been "lapped" and no buffer allocations have occurred recently,
3618 : * or if the bgwriter has been effectively disabled by setting
3619 : * bgwriter_lru_maxpages to 0.)
3620 : */
3621 : bool
3622 24660 : BgBufferSync(WritebackContext *wb_context)
3623 : {
3624 : /* info obtained from freelist.c */
3625 : int strategy_buf_id;
3626 : uint32 strategy_passes;
3627 : uint32 recent_alloc;
3628 :
3629 : /*
3630 : * Information saved between calls so we can determine the strategy
3631 : * point's advance rate and avoid scanning already-cleaned buffers.
3632 : */
3633 : static bool saved_info_valid = false;
3634 : static int prev_strategy_buf_id;
3635 : static uint32 prev_strategy_passes;
3636 : static int next_to_clean;
3637 : static uint32 next_passes;
3638 :
3639 : /* Moving averages of allocation rate and clean-buffer density */
3640 : static float smoothed_alloc = 0;
3641 : static float smoothed_density = 10.0;
3642 :
3643 : /* Potentially these could be tunables, but for now, not */
3644 24660 : float smoothing_samples = 16;
3645 24660 : float scan_whole_pool_milliseconds = 120000.0;
3646 :
3647 : /* Used to compute how far we scan ahead */
3648 : long strategy_delta;
3649 : int bufs_to_lap;
3650 : int bufs_ahead;
3651 : float scans_per_alloc;
3652 : int reusable_buffers_est;
3653 : int upcoming_alloc_est;
3654 : int min_scan_buffers;
3655 :
3656 : /* Variables for the scanning loop proper */
3657 : int num_to_scan;
3658 : int num_written;
3659 : int reusable_buffers;
3660 :
3661 : /* Variables for final smoothed_density update */
3662 : long new_strategy_delta;
3663 : uint32 new_recent_alloc;
3664 :
3665 : /*
3666 : * Find out where the clock-sweep currently is, and how many buffer
3667 : * allocations have happened since our last call.
3668 : */
3669 24660 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3670 :
3671 : /* Report buffer alloc counts to pgstat */
3672 24660 : PendingBgWriterStats.buf_alloc += recent_alloc;
3673 :
3674 : /*
3675 : * If we're not running the LRU scan, just stop after doing the stats
3676 : * stuff. We mark the saved state invalid so that we can recover sanely
3677 : * if LRU scan is turned back on later.
3678 : */
3679 24660 : if (bgwriter_lru_maxpages <= 0)
3680 : {
3681 78 : saved_info_valid = false;
3682 78 : return true;
3683 : }
3684 :
3685 : /*
3686 : * Compute strategy_delta = how many buffers have been scanned by the
3687 : * clock-sweep since last time. If first time through, assume none. Then
3688 : * see if we are still ahead of the clock-sweep, and if so, how many
3689 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3690 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3691 : * behavior when the passes counts wrap around.
3692 : */
3693 24582 : if (saved_info_valid)
3694 : {
3695 23510 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3696 :
3697 23510 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3698 23510 : strategy_delta += (long) passes_delta * NBuffers;
3699 :
3700 : Assert(strategy_delta >= 0);
3701 :
3702 23510 : if ((int32) (next_passes - strategy_passes) > 0)
3703 : {
3704 : /* we're one pass ahead of the strategy point */
3705 4450 : bufs_to_lap = strategy_buf_id - next_to_clean;
3706 : #ifdef BGW_DEBUG
3707 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3708 : next_passes, next_to_clean,
3709 : strategy_passes, strategy_buf_id,
3710 : strategy_delta, bufs_to_lap);
3711 : #endif
3712 : }
3713 19060 : else if (next_passes == strategy_passes &&
3714 14700 : next_to_clean >= strategy_buf_id)
3715 : {
3716 : /* on same pass, but ahead or at least not behind */
3717 13312 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3718 : #ifdef BGW_DEBUG
3719 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3720 : next_passes, next_to_clean,
3721 : strategy_passes, strategy_buf_id,
3722 : strategy_delta, bufs_to_lap);
3723 : #endif
3724 : }
3725 : else
3726 : {
3727 : /*
3728 : * We're behind, so skip forward to the strategy point and start
3729 : * cleaning from there.
3730 : */
3731 : #ifdef BGW_DEBUG
3732 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3733 : next_passes, next_to_clean,
3734 : strategy_passes, strategy_buf_id,
3735 : strategy_delta);
3736 : #endif
3737 5748 : next_to_clean = strategy_buf_id;
3738 5748 : next_passes = strategy_passes;
3739 5748 : bufs_to_lap = NBuffers;
3740 : }
3741 : }
3742 : else
3743 : {
3744 : /*
3745 : * Initializing at startup or after LRU scanning had been off. Always
3746 : * start at the strategy point.
3747 : */
3748 : #ifdef BGW_DEBUG
3749 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3750 : strategy_passes, strategy_buf_id);
3751 : #endif
3752 1072 : strategy_delta = 0;
3753 1072 : next_to_clean = strategy_buf_id;
3754 1072 : next_passes = strategy_passes;
3755 1072 : bufs_to_lap = NBuffers;
3756 : }
3757 :
3758 : /* Update saved info for next time */
3759 24582 : prev_strategy_buf_id = strategy_buf_id;
3760 24582 : prev_strategy_passes = strategy_passes;
3761 24582 : saved_info_valid = true;
3762 :
3763 : /*
3764 : * Compute how many buffers had to be scanned for each new allocation, ie,
3765 : * 1/density of reusable buffers, and track a moving average of that.
3766 : *
3767 : * If the strategy point didn't move, we don't update the density estimate
3768 : */
3769 24582 : if (strategy_delta > 0 && recent_alloc > 0)
3770 : {
3771 12834 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3772 12834 : smoothed_density += (scans_per_alloc - smoothed_density) /
3773 : smoothing_samples;
3774 : }
3775 :
3776 : /*
3777 : * Estimate how many reusable buffers there are between the current
3778 : * strategy point and where we've scanned ahead to, based on the smoothed
3779 : * density estimate.
3780 : */
3781 24582 : bufs_ahead = NBuffers - bufs_to_lap;
3782 24582 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
3783 :
3784 : /*
3785 : * Track a moving average of recent buffer allocations. Here, rather than
3786 : * a true average we want a fast-attack, slow-decline behavior: we
3787 : * immediately follow any increase.
3788 : */
3789 24582 : if (smoothed_alloc <= (float) recent_alloc)
3790 6912 : smoothed_alloc = recent_alloc;
3791 : else
3792 17670 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
3793 : smoothing_samples;
3794 :
3795 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
3796 24582 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
3797 :
3798 : /*
3799 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
3800 : * eventually underflow to zero, and the underflows produce annoying
3801 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
3802 : * zero, there's no point in tracking smaller and smaller values of
3803 : * smoothed_alloc, so just reset it to exactly zero to avoid this
3804 : * syndrome. It will pop back up as soon as recent_alloc increases.
3805 : */
3806 24582 : if (upcoming_alloc_est == 0)
3807 4174 : smoothed_alloc = 0;
3808 :
3809 : /*
3810 : * Even in cases where there's been little or no buffer allocation
3811 : * activity, we want to make a small amount of progress through the buffer
3812 : * cache so that as many reusable buffers as possible are clean after an
3813 : * idle period.
3814 : *
3815 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
3816 : * the BGW will be called during the scan_whole_pool time; slice the
3817 : * buffer pool into that many sections.
3818 : */
3819 24582 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
3820 :
3821 24582 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
3822 : {
3823 : #ifdef BGW_DEBUG
3824 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
3825 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
3826 : #endif
3827 10964 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
3828 : }
3829 :
3830 : /*
3831 : * Now write out dirty reusable buffers, working forward from the
3832 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
3833 : * enough buffers to match our estimate of the next cycle's allocation
3834 : * requirements, or hit the bgwriter_lru_maxpages limit.
3835 : */
3836 :
3837 24582 : num_to_scan = bufs_to_lap;
3838 24582 : num_written = 0;
3839 24582 : reusable_buffers = reusable_buffers_est;
3840 :
3841 : /* Execute the LRU scan */
3842 3519068 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
3843 : {
3844 3494488 : int sync_state = SyncOneBuffer(next_to_clean, true,
3845 : wb_context);
3846 :
3847 3494488 : if (++next_to_clean >= NBuffers)
3848 : {
3849 5382 : next_to_clean = 0;
3850 5382 : next_passes++;
3851 : }
3852 3494488 : num_to_scan--;
3853 :
3854 3494488 : if (sync_state & BUF_WRITTEN)
3855 : {
3856 45568 : reusable_buffers++;
3857 45568 : if (++num_written >= bgwriter_lru_maxpages)
3858 : {
3859 2 : PendingBgWriterStats.maxwritten_clean++;
3860 2 : break;
3861 : }
3862 : }
3863 3448920 : else if (sync_state & BUF_REUSABLE)
3864 2700160 : reusable_buffers++;
3865 : }
3866 :
3867 24582 : PendingBgWriterStats.buf_written_clean += num_written;
3868 :
3869 : #ifdef BGW_DEBUG
3870 : elog(DEBUG1, "bgwriter: recent_alloc=%u smoothed=%.2f delta=%ld ahead=%d density=%.2f reusable_est=%d upcoming_est=%d scanned=%d wrote=%d reusable=%d",
3871 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
3872 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
3873 : bufs_to_lap - num_to_scan,
3874 : num_written,
3875 : reusable_buffers - reusable_buffers_est);
3876 : #endif
3877 :
3878 : /*
3879 : * Consider the above scan as being like a new allocation scan.
3880 : * Characterize its density and update the smoothed one based on it. This
3881 : * effectively halves the moving average period in cases where both the
3882 : * strategy and the background writer are doing some useful scanning,
3883 : * which is helpful because a long memory isn't as desirable on the
3884 : * density estimates.
3885 : */
3886 24582 : new_strategy_delta = bufs_to_lap - num_to_scan;
3887 24582 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
3888 24582 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
3889 : {
3890 19862 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
3891 19862 : smoothed_density += (scans_per_alloc - smoothed_density) /
3892 : smoothing_samples;
3893 :
3894 : #ifdef BGW_DEBUG
3895 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
3896 : new_recent_alloc, new_strategy_delta,
3897 : scans_per_alloc, smoothed_density);
3898 : #endif
3899 : }
3900 :
3901 : /* Return true if OK to hibernate */
3902 24582 : return (bufs_to_lap == 0 && recent_alloc == 0);
3903 : }
3904 :
3905 : /*
3906 : * SyncOneBuffer -- process a single buffer during syncing.
3907 : *
3908 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
3909 : * buffers marked recently used, as these are not replacement candidates.
3910 : *
3911 : * Returns a bitmask containing the following flag bits:
3912 : * BUF_WRITTEN: we wrote the buffer.
3913 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
3914 : * pin count 0 and usage count 0.
3915 : *
3916 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
3917 : * after locking it, but we don't care all that much.)
3918 : */
3919 : static int
3920 4039416 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
3921 : {
3922 4039416 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3923 4039416 : int result = 0;
3924 : uint32 buf_state;
3925 : BufferTag tag;
3926 :
3927 : /* Make sure we can handle the pin */
3928 4039416 : ReservePrivateRefCountEntry();
3929 4039416 : ResourceOwnerEnlarge(CurrentResourceOwner);
3930 :
3931 : /*
3932 : * Check whether buffer needs writing.
3933 : *
3934 : * We can make this check without taking the buffer content lock so long
3935 : * as we mark pages dirty in access methods *before* logging changes with
3936 : * XLogInsert(): if someone marks the buffer dirty just after our check we
3937 : * don't worry because our checkpoint.redo points before log record for
3938 : * upcoming changes and so we are not required to write such dirty buffer.
3939 : */
3940 4039416 : buf_state = LockBufHdr(bufHdr);
3941 :
3942 4039416 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
3943 4030192 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3944 : {
3945 2749668 : result |= BUF_REUSABLE;
3946 : }
3947 1289748 : else if (skip_recently_used)
3948 : {
3949 : /* Caller told us not to write recently-used buffers */
3950 748760 : UnlockBufHdr(bufHdr);
3951 748760 : return result;
3952 : }
3953 :
3954 3290656 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
3955 : {
3956 : /* It's clean, so nothing to do */
3957 2700160 : UnlockBufHdr(bufHdr);
3958 2700160 : return result;
3959 : }
3960 :
3961 : /*
3962 : * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
3963 : * buffer is clean by the time we've locked it.)
3964 : */
3965 590496 : PinBuffer_Locked(bufHdr);
3966 :
3967 590496 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
3968 :
3969 590496 : tag = bufHdr->tag;
3970 :
3971 590496 : UnpinBuffer(bufHdr);
3972 :
3973 : /*
3974 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
3975 : * IOContext will always be IOCONTEXT_NORMAL.
3976 : */
3977 590496 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
3978 :
3979 590496 : return result | BUF_WRITTEN;
3980 : }
3981 :
3982 : /*
3983 : * AtEOXact_Buffers - clean up at end of transaction.
3984 : *
3985 : * As of PostgreSQL 8.0, buffer pins should get released by the
3986 : * ResourceOwner mechanism. This routine is just a debugging
3987 : * cross-check that no pins remain.
3988 : */
3989 : void
3990 1121842 : AtEOXact_Buffers(bool isCommit)
3991 : {
3992 1121842 : CheckForBufferLeaks();
3993 :
3994 1121842 : AtEOXact_LocalBuffers(isCommit);
3995 :
3996 : Assert(PrivateRefCountOverflowed == 0);
3997 1121842 : }
3998 :
3999 : /*
4000 : * Initialize access to shared buffer pool
4001 : *
4002 : * This is called during backend startup (whether standalone or under the
4003 : * postmaster). It sets up for this backend's access to the already-existing
4004 : * buffer pool.
4005 : */
4006 : void
4007 45016 : InitBufferManagerAccess(void)
4008 : {
4009 : HASHCTL hash_ctl;
4010 :
4011 : /*
4012 : * An advisory limit on the number of pins each backend should hold, based
4013 : * on shared_buffers and the maximum number of connections possible.
4014 : * That's very pessimistic, but outside toy-sized shared_buffers it should
4015 : * allow plenty of pins. LimitAdditionalPins() and
4016 : * GetAdditionalPinLimit() can be used to check the remaining balance.
4017 : */
4018 45016 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4019 :
4020 45016 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4021 :
4022 45016 : hash_ctl.keysize = sizeof(int32);
4023 45016 : hash_ctl.entrysize = sizeof(PrivateRefCountEntry);
4024 :
4025 45016 : PrivateRefCountHash = hash_create("PrivateRefCount", 100, &hash_ctl,
4026 : HASH_ELEM | HASH_BLOBS);
4027 :
4028 : /*
4029 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4030 : * the corresponding phase of backend shutdown.
4031 : */
4032 : Assert(MyProc != NULL);
4033 45016 : on_shmem_exit(AtProcExit_Buffers, 0);
4034 45016 : }
4035 :
4036 : /*
4037 : * During backend exit, ensure that we released all shared-buffer locks and
4038 : * assert that we have no remaining pins.
4039 : */
4040 : static void
4041 45016 : AtProcExit_Buffers(int code, Datum arg)
4042 : {
4043 45016 : UnlockBuffers();
4044 :
4045 45016 : CheckForBufferLeaks();
4046 :
4047 : /* localbuf.c needs a chance too */
4048 45016 : AtProcExit_LocalBuffers();
4049 45016 : }
4050 :
4051 : /*
4052 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4053 : *
4054 : * As of PostgreSQL 8.0, buffer pins should get released by the
4055 : * ResourceOwner mechanism. This routine is just a debugging
4056 : * cross-check that no pins remain.
4057 : */
4058 : static void
4059 1166858 : CheckForBufferLeaks(void)
4060 : {
4061 : #ifdef USE_ASSERT_CHECKING
4062 : int RefCountErrors = 0;
4063 : PrivateRefCountEntry *res;
4064 : int i;
4065 : char *s;
4066 :
4067 : /* check the array */
4068 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4069 : {
4070 : res = &PrivateRefCountArray[i];
4071 :
4072 : if (res->buffer != InvalidBuffer)
4073 : {
4074 : s = DebugPrintBufferRefcount(res->buffer);
4075 : elog(WARNING, "buffer refcount leak: %s", s);
4076 : pfree(s);
4077 :
4078 : RefCountErrors++;
4079 : }
4080 : }
4081 :
4082 : /* if necessary search the hash */
4083 : if (PrivateRefCountOverflowed)
4084 : {
4085 : HASH_SEQ_STATUS hstat;
4086 :
4087 : hash_seq_init(&hstat, PrivateRefCountHash);
4088 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
4089 : {
4090 : s = DebugPrintBufferRefcount(res->buffer);
4091 : elog(WARNING, "buffer refcount leak: %s", s);
4092 : pfree(s);
4093 : RefCountErrors++;
4094 : }
4095 : }
4096 :
4097 : Assert(RefCountErrors == 0);
4098 : #endif
4099 1166858 : }
4100 :
4101 : #ifdef USE_ASSERT_CHECKING
4102 : /*
4103 : * Check for exclusive-locked catalog buffers. This is the core of
4104 : * AssertCouldGetRelation().
4105 : *
4106 : * A backend would self-deadlock on LWLocks if the catalog scan read the
4107 : * exclusive-locked buffer. The main threat is exclusive-locked buffers of
4108 : * catalogs used in relcache, because a catcache search on any catalog may
4109 : * build that catalog's relcache entry. We don't have an inventory of
4110 : * catalogs relcache uses, so just check buffers of most catalogs.
4111 : *
4112 : * It's better to minimize waits while holding an exclusive buffer lock, so it
4113 : * would be nice to broaden this check not to be catalog-specific. However,
4114 : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4115 : * read tables. That is deadlock-free as long as there's no loop in the
4116 : * dependency graph: modifying table A may cause an opclass to read table B,
4117 : * but it must not cause a read of table A.
4118 : */
4119 : void
4120 : AssertBufferLocksPermitCatalogRead(void)
4121 : {
4122 : ForEachLWLockHeldByMe(AssertNotCatalogBufferLock, NULL);
4123 : }
4124 :
4125 : static void
4126 : AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
4127 : void *unused_context)
4128 : {
4129 : BufferDesc *bufHdr;
4130 : BufferTag tag;
4131 : Oid relid;
4132 :
4133 : if (mode != LW_EXCLUSIVE)
4134 : return;
4135 :
4136 : if (!((BufferDescPadded *) lock > BufferDescriptors &&
4137 : (BufferDescPadded *) lock < BufferDescriptors + NBuffers))
4138 : return; /* not a buffer lock */
4139 :
4140 : bufHdr = (BufferDesc *)
4141 : ((char *) lock - offsetof(BufferDesc, content_lock));
4142 : tag = bufHdr->tag;
4143 :
4144 : /*
4145 : * This relNumber==relid assumption holds until a catalog experiences
4146 : * VACUUM FULL or similar. After a command like that, relNumber will be
4147 : * in the normal (non-catalog) range, and we lose the ability to detect
4148 : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4149 : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4150 : * held lock.
4151 : */
4152 : relid = tag.relNumber;
4153 :
4154 : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4155 : return;
4156 :
4157 : Assert(!IsCatalogRelationOid(relid));
4158 : }
4159 : #endif
4160 :
4161 :
4162 : /*
4163 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4164 : */
4165 : char *
4166 80 : DebugPrintBufferRefcount(Buffer buffer)
4167 : {
4168 : BufferDesc *buf;
4169 : int32 loccount;
4170 : char *result;
4171 : ProcNumber backend;
4172 : uint32 buf_state;
4173 :
4174 : Assert(BufferIsValid(buffer));
4175 80 : if (BufferIsLocal(buffer))
4176 : {
4177 32 : buf = GetLocalBufferDescriptor(-buffer - 1);
4178 32 : loccount = LocalRefCount[-buffer - 1];
4179 32 : backend = MyProcNumber;
4180 : }
4181 : else
4182 : {
4183 48 : buf = GetBufferDescriptor(buffer - 1);
4184 48 : loccount = GetPrivateRefCount(buffer);
4185 48 : backend = INVALID_PROC_NUMBER;
4186 : }
4187 :
4188 : /* theoretically we should lock the bufhdr here */
4189 80 : buf_state = pg_atomic_read_u32(&buf->state);
4190 :
4191 80 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%x, refcount=%u %d)",
4192 : buffer,
4193 80 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4194 : BufTagGetForkNum(&buf->tag)).str,
4195 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4196 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4197 80 : return result;
4198 : }
4199 :
4200 : /*
4201 : * CheckPointBuffers
4202 : *
4203 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4204 : *
4205 : * Note: temporary relations do not participate in checkpoints, so they don't
4206 : * need to be flushed.
4207 : */
4208 : void
4209 3466 : CheckPointBuffers(int flags)
4210 : {
4211 3466 : BufferSync(flags);
4212 3466 : }
4213 :
4214 : /*
4215 : * BufferGetBlockNumber
4216 : * Returns the block number associated with a buffer.
4217 : *
4218 : * Note:
4219 : * Assumes that the buffer is valid and pinned, else the
4220 : * value may be obsolete immediately...
4221 : */
4222 : BlockNumber
4223 100936396 : BufferGetBlockNumber(Buffer buffer)
4224 : {
4225 : BufferDesc *bufHdr;
4226 :
4227 : Assert(BufferIsPinned(buffer));
4228 :
4229 100936396 : if (BufferIsLocal(buffer))
4230 3809030 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4231 : else
4232 97127366 : bufHdr = GetBufferDescriptor(buffer - 1);
4233 :
4234 : /* pinned, so OK to read tag without spinlock */
4235 100936396 : return bufHdr->tag.blockNum;
4236 : }
4237 :
4238 : /*
4239 : * BufferGetTag
4240 : * Returns the relfilelocator, fork number and block number associated with
4241 : * a buffer.
4242 : */
4243 : void
4244 30152996 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4245 : BlockNumber *blknum)
4246 : {
4247 : BufferDesc *bufHdr;
4248 :
4249 : /* Do the same checks as BufferGetBlockNumber. */
4250 : Assert(BufferIsPinned(buffer));
4251 :
4252 30152996 : if (BufferIsLocal(buffer))
4253 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4254 : else
4255 30152996 : bufHdr = GetBufferDescriptor(buffer - 1);
4256 :
4257 : /* pinned, so OK to read tag without spinlock */
4258 30152996 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4259 30152996 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4260 30152996 : *blknum = bufHdr->tag.blockNum;
4261 30152996 : }
4262 :
4263 : /*
4264 : * FlushBuffer
4265 : * Physically write out a shared buffer.
4266 : *
4267 : * NOTE: this actually just passes the buffer contents to the kernel; the
4268 : * real write to disk won't happen until the kernel feels like it. This
4269 : * is okay from our point of view since we can redo the changes from WAL.
4270 : * However, we will need to force the changes to disk via fsync before
4271 : * we can checkpoint WAL.
4272 : *
4273 : * The caller must hold a pin on the buffer and have share-locked the
4274 : * buffer contents. (Note: a share-lock does not prevent updates of
4275 : * hint bits in the buffer, so the page could change while the write
4276 : * is in progress, but we assume that that will not invalidate the data
4277 : * written.)
4278 : *
4279 : * If the caller has an smgr reference for the buffer's relation, pass it
4280 : * as the second parameter. If not, pass NULL.
4281 : */
4282 : static void
4283 1122730 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4284 : IOContext io_context)
4285 : {
4286 : XLogRecPtr recptr;
4287 : ErrorContextCallback errcallback;
4288 : instr_time io_start;
4289 : Block bufBlock;
4290 : char *bufToWrite;
4291 : uint32 buf_state;
4292 :
4293 : /*
4294 : * Try to start an I/O operation. If StartBufferIO returns false, then
4295 : * someone else flushed the buffer before we could, so we need not do
4296 : * anything.
4297 : */
4298 1122730 : if (!StartBufferIO(buf, false, false))
4299 26 : return;
4300 :
4301 : /* Setup error traceback support for ereport() */
4302 1122704 : errcallback.callback = shared_buffer_write_error_callback;
4303 1122704 : errcallback.arg = buf;
4304 1122704 : errcallback.previous = error_context_stack;
4305 1122704 : error_context_stack = &errcallback;
4306 :
4307 : /* Find smgr relation for buffer */
4308 1122704 : if (reln == NULL)
4309 1114430 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4310 :
4311 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4312 : buf->tag.blockNum,
4313 : reln->smgr_rlocator.locator.spcOid,
4314 : reln->smgr_rlocator.locator.dbOid,
4315 : reln->smgr_rlocator.locator.relNumber);
4316 :
4317 1122704 : buf_state = LockBufHdr(buf);
4318 :
4319 : /*
4320 : * Run PageGetLSN while holding header lock, since we don't have the
4321 : * buffer locked exclusively in all cases.
4322 : */
4323 1122704 : recptr = BufferGetLSN(buf);
4324 :
4325 : /* To check if block content changes while flushing. - vadim 01/17/97 */
4326 1122704 : UnlockBufHdrExt(buf, buf_state,
4327 : 0, BM_JUST_DIRTIED,
4328 : 0);
4329 :
4330 : /*
4331 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4332 : * rule that log updates must hit disk before any of the data-file changes
4333 : * they describe do.
4334 : *
4335 : * However, this rule does not apply to unlogged relations, which will be
4336 : * lost after a crash anyway. Most unlogged relation pages do not bear
4337 : * LSNs since we never emit WAL records for them, and therefore flushing
4338 : * up through the buffer LSN would be useless, but harmless. However,
4339 : * GiST indexes use LSNs internally to track page-splits, and therefore
4340 : * unlogged GiST pages bear "fake" LSNs generated by
4341 : * GetFakeLSNForUnloggedRel. It is unlikely but possible that the fake
4342 : * LSN counter could advance past the WAL insertion point; and if it did
4343 : * happen, attempting to flush WAL through that location would fail, with
4344 : * disastrous system-wide consequences. To make sure that can't happen,
4345 : * skip the flush if the buffer isn't permanent.
4346 : */
4347 1122704 : if (buf_state & BM_PERMANENT)
4348 1119102 : XLogFlush(recptr);
4349 :
4350 : /*
4351 : * Now it's safe to write the buffer to disk. Note that no one else should
4352 : * have been able to write it, while we were busy with log flushing,
4353 : * because we got the exclusive right to perform I/O by setting the
4354 : * BM_IO_IN_PROGRESS bit.
4355 : */
4356 1122704 : bufBlock = BufHdrGetBlock(buf);
4357 :
4358 : /*
4359 : * Update page checksum if desired. Since we have only shared lock on the
4360 : * buffer, other processes might be updating hint bits in it, so we must
4361 : * copy the page to private storage if we do checksumming.
4362 : */
4363 1122704 : bufToWrite = PageSetChecksumCopy((Page) bufBlock, buf->tag.blockNum);
4364 :
4365 1122704 : io_start = pgstat_prepare_io_time(track_io_timing);
4366 :
4367 : /*
4368 : * bufToWrite is either the shared buffer or a copy, as appropriate.
4369 : */
4370 1122704 : smgrwrite(reln,
4371 1122704 : BufTagGetForkNum(&buf->tag),
4372 : buf->tag.blockNum,
4373 : bufToWrite,
4374 : false);
4375 :
4376 : /*
4377 : * When a strategy is in use, only flushes of dirty buffers already in the
4378 : * strategy ring are counted as strategy writes (IOCONTEXT
4379 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4380 : * statistics tracking.
4381 : *
4382 : * If a shared buffer initially added to the ring must be flushed before
4383 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4384 : *
4385 : * If a shared buffer which was added to the ring later because the
4386 : * current strategy buffer is pinned or in use or because all strategy
4387 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4388 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4389 : * (from_ring will be false).
4390 : *
4391 : * When a strategy is not in use, the write can only be a "regular" write
4392 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4393 : */
4394 1122704 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
4395 : IOOP_WRITE, io_start, 1, BLCKSZ);
4396 :
4397 1122704 : pgBufferUsage.shared_blks_written++;
4398 :
4399 : /*
4400 : * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
4401 : * end the BM_IO_IN_PROGRESS state.
4402 : */
4403 1122704 : TerminateBufferIO(buf, true, 0, true, false);
4404 :
4405 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4406 : buf->tag.blockNum,
4407 : reln->smgr_rlocator.locator.spcOid,
4408 : reln->smgr_rlocator.locator.dbOid,
4409 : reln->smgr_rlocator.locator.relNumber);
4410 :
4411 : /* Pop the error context stack */
4412 1122704 : error_context_stack = errcallback.previous;
4413 : }
4414 :
4415 : /*
4416 : * Convenience wrapper around FlushBuffer() that locks/unlocks the buffer
4417 : * before/after calling FlushBuffer().
4418 : */
4419 : static void
4420 600742 : FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
4421 : IOObject io_object, IOContext io_context)
4422 : {
4423 600742 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
4424 600742 : FlushBuffer(buf, reln, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4425 600742 : LWLockRelease(BufferDescriptorGetContentLock(buf));
4426 600742 : }
4427 :
4428 : /*
4429 : * RelationGetNumberOfBlocksInFork
4430 : * Determines the current number of pages in the specified relation fork.
4431 : *
4432 : * Note that the accuracy of the result will depend on the details of the
4433 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4434 : * it might not be.
4435 : */
4436 : BlockNumber
4437 3769470 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4438 : {
4439 3769470 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4440 : {
4441 : /*
4442 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4443 : * tableam returns the size in bytes - but for the purpose of this
4444 : * routine, we want the number of blocks. Therefore divide, rounding
4445 : * up.
4446 : */
4447 : uint64 szbytes;
4448 :
4449 2867738 : szbytes = table_relation_size(relation, forkNum);
4450 :
4451 2867700 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4452 : }
4453 901732 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4454 : {
4455 901732 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4456 : }
4457 : else
4458 : Assert(false);
4459 :
4460 0 : return 0; /* keep compiler quiet */
4461 : }
4462 :
4463 : /*
4464 : * BufferIsPermanent
4465 : * Determines whether a buffer will potentially still be around after
4466 : * a crash. Caller must hold a buffer pin.
4467 : */
4468 : bool
4469 19045596 : BufferIsPermanent(Buffer buffer)
4470 : {
4471 : BufferDesc *bufHdr;
4472 :
4473 : /* Local buffers are used only for temp relations. */
4474 19045596 : if (BufferIsLocal(buffer))
4475 1253940 : return false;
4476 :
4477 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4478 : Assert(BufferIsValid(buffer));
4479 : Assert(BufferIsPinned(buffer));
4480 :
4481 : /*
4482 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4483 : * need not bother with the buffer header spinlock. Even if someone else
4484 : * changes the buffer header state while we're doing this, the state is
4485 : * changed atomically, so we'll read the old value or the new value, but
4486 : * not random garbage.
4487 : */
4488 17791656 : bufHdr = GetBufferDescriptor(buffer - 1);
4489 17791656 : return (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT) != 0;
4490 : }
4491 :
4492 : /*
4493 : * BufferGetLSNAtomic
4494 : * Retrieves the LSN of the buffer atomically using a buffer header lock.
4495 : * This is necessary for some callers who may not have an exclusive lock
4496 : * on the buffer.
4497 : */
4498 : XLogRecPtr
4499 14323544 : BufferGetLSNAtomic(Buffer buffer)
4500 : {
4501 14323544 : char *page = BufferGetPage(buffer);
4502 : BufferDesc *bufHdr;
4503 : XLogRecPtr lsn;
4504 :
4505 : /*
4506 : * If we don't need locking for correctness, fastpath out.
4507 : */
4508 14323544 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4509 481030 : return PageGetLSN(page);
4510 :
4511 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4512 : Assert(BufferIsValid(buffer));
4513 : Assert(BufferIsPinned(buffer));
4514 :
4515 13842514 : bufHdr = GetBufferDescriptor(buffer - 1);
4516 13842514 : LockBufHdr(bufHdr);
4517 13842514 : lsn = PageGetLSN(page);
4518 13842514 : UnlockBufHdr(bufHdr);
4519 :
4520 13842514 : return lsn;
4521 : }
4522 :
4523 : /* ---------------------------------------------------------------------
4524 : * DropRelationBuffers
4525 : *
4526 : * This function removes from the buffer pool all the pages of the
4527 : * specified relation forks that have block numbers >= firstDelBlock.
4528 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4529 : * Dirty pages are simply dropped, without bothering to write them
4530 : * out first. Therefore, this is NOT rollback-able, and so should be
4531 : * used only with extreme caution!
4532 : *
4533 : * Currently, this is called only from smgr.c when the underlying file
4534 : * is about to be deleted or truncated (firstDelBlock is needed for
4535 : * the truncation case). The data in the affected pages would therefore
4536 : * be deleted momentarily anyway, and there is no point in writing it.
4537 : * It is the responsibility of higher-level code to ensure that the
4538 : * deletion or truncation does not lose any data that could be needed
4539 : * later. It is also the responsibility of higher-level code to ensure
4540 : * that no other process could be trying to load more pages of the
4541 : * relation into buffers.
4542 : * --------------------------------------------------------------------
4543 : */
4544 : void
4545 1268 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4546 : int nforks, BlockNumber *firstDelBlock)
4547 : {
4548 : int i;
4549 : int j;
4550 : RelFileLocatorBackend rlocator;
4551 : BlockNumber nForkBlock[MAX_FORKNUM];
4552 1268 : uint64 nBlocksToInvalidate = 0;
4553 :
4554 1268 : rlocator = smgr_reln->smgr_rlocator;
4555 :
4556 : /* If it's a local relation, it's localbuf.c's problem. */
4557 1268 : if (RelFileLocatorBackendIsTemp(rlocator))
4558 : {
4559 748 : if (rlocator.backend == MyProcNumber)
4560 748 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4561 : firstDelBlock);
4562 :
4563 830 : return;
4564 : }
4565 :
4566 : /*
4567 : * To remove all the pages of the specified relation forks from the buffer
4568 : * pool, we need to scan the entire buffer pool but we can optimize it by
4569 : * finding the buffers from BufMapping table provided we know the exact
4570 : * size of each fork of the relation. The exact size is required to ensure
4571 : * that we don't leave any buffer for the relation being dropped as
4572 : * otherwise the background writer or checkpointer can lead to a PANIC
4573 : * error while flushing buffers corresponding to files that don't exist.
4574 : *
4575 : * To know the exact size, we rely on the size cached for each fork by us
4576 : * during recovery which limits the optimization to recovery and on
4577 : * standbys but we can easily extend it once we have shared cache for
4578 : * relation size.
4579 : *
4580 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4581 : * and the future writes keeps the cached value up-to-date. See
4582 : * smgrextend. It is possible that the value of the first lseek is smaller
4583 : * than the actual number of existing blocks in the file due to buggy
4584 : * Linux kernels that might not have accounted for the recent write. But
4585 : * that should be fine because there must not be any buffers after that
4586 : * file size.
4587 : */
4588 720 : for (i = 0; i < nforks; i++)
4589 : {
4590 : /* Get the number of blocks for a relation's fork */
4591 612 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4592 :
4593 612 : if (nForkBlock[i] == InvalidBlockNumber)
4594 : {
4595 412 : nBlocksToInvalidate = InvalidBlockNumber;
4596 412 : break;
4597 : }
4598 :
4599 : /* calculate the number of blocks to be invalidated */
4600 200 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4601 : }
4602 :
4603 : /*
4604 : * We apply the optimization iff the total number of blocks to invalidate
4605 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4606 : */
4607 520 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4608 108 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4609 : {
4610 222 : for (j = 0; j < nforks; j++)
4611 140 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4612 140 : nForkBlock[j], firstDelBlock[j]);
4613 82 : return;
4614 : }
4615 :
4616 5616054 : for (i = 0; i < NBuffers; i++)
4617 : {
4618 5615616 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4619 :
4620 : /*
4621 : * We can make this a tad faster by prechecking the buffer tag before
4622 : * we attempt to lock the buffer; this saves a lot of lock
4623 : * acquisitions in typical cases. It should be safe because the
4624 : * caller must have AccessExclusiveLock on the relation, or some other
4625 : * reason to be certain that no one is loading new pages of the rel
4626 : * into the buffer pool. (Otherwise we might well miss such pages
4627 : * entirely.) Therefore, while the tag might be changing while we
4628 : * look at it, it can't be changing *to* a value we care about, only
4629 : * *away* from such a value. So false negatives are impossible, and
4630 : * false positives are safe because we'll recheck after getting the
4631 : * buffer lock.
4632 : *
4633 : * We could check forkNum and blockNum as well as the rlocator, but
4634 : * the incremental win from doing so seems small.
4635 : */
4636 5615616 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4637 5601736 : continue;
4638 :
4639 13880 : LockBufHdr(bufHdr);
4640 :
4641 35634 : for (j = 0; j < nforks; j++)
4642 : {
4643 25010 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4644 25010 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4645 13686 : bufHdr->tag.blockNum >= firstDelBlock[j])
4646 : {
4647 3256 : InvalidateBuffer(bufHdr); /* releases spinlock */
4648 3256 : break;
4649 : }
4650 : }
4651 13880 : if (j >= nforks)
4652 10624 : UnlockBufHdr(bufHdr);
4653 : }
4654 : }
4655 :
4656 : /* ---------------------------------------------------------------------
4657 : * DropRelationsAllBuffers
4658 : *
4659 : * This function removes from the buffer pool all the pages of all
4660 : * forks of the specified relations. It's equivalent to calling
4661 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4662 : * --------------------------------------------------------------------
4663 : */
4664 : void
4665 27508 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4666 : {
4667 : int i;
4668 27508 : int n = 0;
4669 : SMgrRelation *rels;
4670 : BlockNumber (*block)[MAX_FORKNUM + 1];
4671 27508 : uint64 nBlocksToInvalidate = 0;
4672 : RelFileLocator *locators;
4673 27508 : bool cached = true;
4674 : bool use_bsearch;
4675 :
4676 27508 : if (nlocators == 0)
4677 0 : return;
4678 :
4679 27508 : rels = palloc(sizeof(SMgrRelation) * nlocators); /* non-local relations */
4680 :
4681 : /* If it's a local relation, it's localbuf.c's problem. */
4682 119938 : for (i = 0; i < nlocators; i++)
4683 : {
4684 92430 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4685 : {
4686 6358 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4687 6358 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4688 : }
4689 : else
4690 86072 : rels[n++] = smgr_reln[i];
4691 : }
4692 :
4693 : /*
4694 : * If there are no non-local relations, then we're done. Release the
4695 : * memory and return.
4696 : */
4697 27508 : if (n == 0)
4698 : {
4699 1656 : pfree(rels);
4700 1656 : return;
4701 : }
4702 :
4703 : /*
4704 : * This is used to remember the number of blocks for all the relations
4705 : * forks.
4706 : */
4707 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4708 25852 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4709 :
4710 : /*
4711 : * We can avoid scanning the entire buffer pool if we know the exact size
4712 : * of each of the given relation forks. See DropRelationBuffers.
4713 : */
4714 54052 : for (i = 0; i < n && cached; i++)
4715 : {
4716 43986 : for (int j = 0; j <= MAX_FORKNUM; j++)
4717 : {
4718 : /* Get the number of blocks for a relation's fork. */
4719 40070 : block[i][j] = smgrnblocks_cached(rels[i], j);
4720 :
4721 : /* We need to only consider the relation forks that exists. */
4722 40070 : if (block[i][j] == InvalidBlockNumber)
4723 : {
4724 35820 : if (!smgrexists(rels[i], j))
4725 11536 : continue;
4726 24284 : cached = false;
4727 24284 : break;
4728 : }
4729 :
4730 : /* calculate the total number of blocks to be invalidated */
4731 4250 : nBlocksToInvalidate += block[i][j];
4732 : }
4733 : }
4734 :
4735 : /*
4736 : * We apply the optimization iff the total number of blocks to invalidate
4737 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4738 : */
4739 25852 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4740 : {
4741 2674 : for (i = 0; i < n; i++)
4742 : {
4743 7340 : for (int j = 0; j <= MAX_FORKNUM; j++)
4744 : {
4745 : /* ignore relation forks that doesn't exist */
4746 5872 : if (!BlockNumberIsValid(block[i][j]))
4747 4386 : continue;
4748 :
4749 : /* drop all the buffers for a particular relation fork */
4750 1486 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4751 1486 : j, block[i][j], 0);
4752 : }
4753 : }
4754 :
4755 1206 : pfree(block);
4756 1206 : pfree(rels);
4757 1206 : return;
4758 : }
4759 :
4760 24646 : pfree(block);
4761 24646 : locators = palloc(sizeof(RelFileLocator) * n); /* non-local relations */
4762 109250 : for (i = 0; i < n; i++)
4763 84604 : locators[i] = rels[i]->smgr_rlocator.locator;
4764 :
4765 : /*
4766 : * For low number of relations to drop just use a simple walk through, to
4767 : * save the bsearch overhead. The threshold to use is rather a guess than
4768 : * an exactly determined value, as it depends on many factors (CPU and RAM
4769 : * speeds, amount of shared buffers etc.).
4770 : */
4771 24646 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
4772 :
4773 : /* sort the list of rlocators if necessary */
4774 24646 : if (use_bsearch)
4775 346 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
4776 :
4777 267079238 : for (i = 0; i < NBuffers; i++)
4778 : {
4779 267054592 : RelFileLocator *rlocator = NULL;
4780 267054592 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4781 :
4782 : /*
4783 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4784 : * saves some cycles.
4785 : */
4786 :
4787 267054592 : if (!use_bsearch)
4788 : {
4789 : int j;
4790 :
4791 1061607896 : for (j = 0; j < n; j++)
4792 : {
4793 798449026 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
4794 : {
4795 177578 : rlocator = &locators[j];
4796 177578 : break;
4797 : }
4798 : }
4799 : }
4800 : else
4801 : {
4802 : RelFileLocator locator;
4803 :
4804 3718144 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
4805 3718144 : rlocator = bsearch(&locator,
4806 : locators, n, sizeof(RelFileLocator),
4807 : rlocator_comparator);
4808 : }
4809 :
4810 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
4811 267054592 : if (rlocator == NULL)
4812 266873436 : continue;
4813 :
4814 181156 : LockBufHdr(bufHdr);
4815 181156 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
4816 181156 : InvalidateBuffer(bufHdr); /* releases spinlock */
4817 : else
4818 0 : UnlockBufHdr(bufHdr);
4819 : }
4820 :
4821 24646 : pfree(locators);
4822 24646 : pfree(rels);
4823 : }
4824 :
4825 : /* ---------------------------------------------------------------------
4826 : * FindAndDropRelationBuffers
4827 : *
4828 : * This function performs look up in BufMapping table and removes from the
4829 : * buffer pool all the pages of the specified relation fork that has block
4830 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
4831 : * pages are removed.)
4832 : * --------------------------------------------------------------------
4833 : */
4834 : static void
4835 1626 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
4836 : BlockNumber nForkBlock,
4837 : BlockNumber firstDelBlock)
4838 : {
4839 : BlockNumber curBlock;
4840 :
4841 3918 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
4842 : {
4843 : uint32 bufHash; /* hash value for tag */
4844 : BufferTag bufTag; /* identity of requested block */
4845 : LWLock *bufPartitionLock; /* buffer partition lock for it */
4846 : int buf_id;
4847 : BufferDesc *bufHdr;
4848 :
4849 : /* create a tag so we can lookup the buffer */
4850 2292 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
4851 :
4852 : /* determine its hash code and partition lock ID */
4853 2292 : bufHash = BufTableHashCode(&bufTag);
4854 2292 : bufPartitionLock = BufMappingPartitionLock(bufHash);
4855 :
4856 : /* Check that it is in the buffer pool. If not, do nothing. */
4857 2292 : LWLockAcquire(bufPartitionLock, LW_SHARED);
4858 2292 : buf_id = BufTableLookup(&bufTag, bufHash);
4859 2292 : LWLockRelease(bufPartitionLock);
4860 :
4861 2292 : if (buf_id < 0)
4862 236 : continue;
4863 :
4864 2056 : bufHdr = GetBufferDescriptor(buf_id);
4865 :
4866 : /*
4867 : * We need to lock the buffer header and recheck if the buffer is
4868 : * still associated with the same block because the buffer could be
4869 : * evicted by some other backend loading blocks for a different
4870 : * relation after we release lock on the BufMapping table.
4871 : */
4872 2056 : LockBufHdr(bufHdr);
4873 :
4874 4112 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
4875 2056 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
4876 2056 : bufHdr->tag.blockNum >= firstDelBlock)
4877 2056 : InvalidateBuffer(bufHdr); /* releases spinlock */
4878 : else
4879 0 : UnlockBufHdr(bufHdr);
4880 : }
4881 1626 : }
4882 :
4883 : /* ---------------------------------------------------------------------
4884 : * DropDatabaseBuffers
4885 : *
4886 : * This function removes all the buffers in the buffer cache for a
4887 : * particular database. Dirty pages are simply dropped, without
4888 : * bothering to write them out first. This is used when we destroy a
4889 : * database, to avoid trying to flush data to disk when the directory
4890 : * tree no longer exists. Implementation is pretty similar to
4891 : * DropRelationBuffers() which is for destroying just one relation.
4892 : * --------------------------------------------------------------------
4893 : */
4894 : void
4895 144 : DropDatabaseBuffers(Oid dbid)
4896 : {
4897 : int i;
4898 :
4899 : /*
4900 : * We needn't consider local buffers, since by assumption the target
4901 : * database isn't our own.
4902 : */
4903 :
4904 961424 : for (i = 0; i < NBuffers; i++)
4905 : {
4906 961280 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4907 :
4908 : /*
4909 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4910 : * saves some cycles.
4911 : */
4912 961280 : if (bufHdr->tag.dbOid != dbid)
4913 936414 : continue;
4914 :
4915 24866 : LockBufHdr(bufHdr);
4916 24866 : if (bufHdr->tag.dbOid == dbid)
4917 24866 : InvalidateBuffer(bufHdr); /* releases spinlock */
4918 : else
4919 0 : UnlockBufHdr(bufHdr);
4920 : }
4921 144 : }
4922 :
4923 : /* ---------------------------------------------------------------------
4924 : * FlushRelationBuffers
4925 : *
4926 : * This function writes all dirty pages of a relation out to disk
4927 : * (or more accurately, out to kernel disk buffers), ensuring that the
4928 : * kernel has an up-to-date view of the relation.
4929 : *
4930 : * Generally, the caller should be holding AccessExclusiveLock on the
4931 : * target relation to ensure that no other backend is busy dirtying
4932 : * more blocks of the relation; the effects can't be expected to last
4933 : * after the lock is released.
4934 : *
4935 : * XXX currently it sequentially searches the buffer pool, should be
4936 : * changed to more clever ways of searching. This routine is not
4937 : * used in any performance-critical code paths, so it's not worth
4938 : * adding additional overhead to normal paths to make it go faster.
4939 : * --------------------------------------------------------------------
4940 : */
4941 : void
4942 276 : FlushRelationBuffers(Relation rel)
4943 : {
4944 : int i;
4945 : BufferDesc *bufHdr;
4946 276 : SMgrRelation srel = RelationGetSmgr(rel);
4947 :
4948 276 : if (RelationUsesLocalBuffers(rel))
4949 : {
4950 1818 : for (i = 0; i < NLocBuffer; i++)
4951 : {
4952 : uint32 buf_state;
4953 :
4954 1800 : bufHdr = GetLocalBufferDescriptor(i);
4955 1800 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
4956 600 : ((buf_state = pg_atomic_read_u32(&bufHdr->state)) &
4957 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4958 : {
4959 : ErrorContextCallback errcallback;
4960 :
4961 : /* Setup error traceback support for ereport() */
4962 600 : errcallback.callback = local_buffer_write_error_callback;
4963 600 : errcallback.arg = bufHdr;
4964 600 : errcallback.previous = error_context_stack;
4965 600 : error_context_stack = &errcallback;
4966 :
4967 : /* Make sure we can handle the pin */
4968 600 : ReservePrivateRefCountEntry();
4969 600 : ResourceOwnerEnlarge(CurrentResourceOwner);
4970 :
4971 : /*
4972 : * Pin/unpin mostly to make valgrind work, but it also seems
4973 : * like the right thing to do.
4974 : */
4975 600 : PinLocalBuffer(bufHdr, false);
4976 :
4977 :
4978 600 : FlushLocalBuffer(bufHdr, srel);
4979 :
4980 600 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
4981 :
4982 : /* Pop the error context stack */
4983 600 : error_context_stack = errcallback.previous;
4984 : }
4985 : }
4986 :
4987 18 : return;
4988 : }
4989 :
4990 3024386 : for (i = 0; i < NBuffers; i++)
4991 : {
4992 : uint32 buf_state;
4993 :
4994 3024128 : bufHdr = GetBufferDescriptor(i);
4995 :
4996 : /*
4997 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4998 : * saves some cycles.
4999 : */
5000 3024128 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
5001 3023686 : continue;
5002 :
5003 : /* Make sure we can handle the pin */
5004 442 : ReservePrivateRefCountEntry();
5005 442 : ResourceOwnerEnlarge(CurrentResourceOwner);
5006 :
5007 442 : buf_state = LockBufHdr(bufHdr);
5008 442 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5009 442 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5010 : {
5011 356 : PinBuffer_Locked(bufHdr);
5012 356 : FlushUnlockedBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5013 356 : UnpinBuffer(bufHdr);
5014 : }
5015 : else
5016 86 : UnlockBufHdr(bufHdr);
5017 : }
5018 : }
5019 :
5020 : /* ---------------------------------------------------------------------
5021 : * FlushRelationsAllBuffers
5022 : *
5023 : * This function flushes out of the buffer pool all the pages of all
5024 : * forks of the specified smgr relations. It's equivalent to calling
5025 : * FlushRelationBuffers once per relation. The relations are assumed not
5026 : * to use local buffers.
5027 : * --------------------------------------------------------------------
5028 : */
5029 : void
5030 26 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
5031 : {
5032 : int i;
5033 : SMgrSortArray *srels;
5034 : bool use_bsearch;
5035 :
5036 26 : if (nrels == 0)
5037 0 : return;
5038 :
5039 : /* fill-in array for qsort */
5040 26 : srels = palloc(sizeof(SMgrSortArray) * nrels);
5041 :
5042 54 : for (i = 0; i < nrels; i++)
5043 : {
5044 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5045 :
5046 28 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5047 28 : srels[i].srel = smgrs[i];
5048 : }
5049 :
5050 : /*
5051 : * Save the bsearch overhead for low number of relations to sync. See
5052 : * DropRelationsAllBuffers for details.
5053 : */
5054 26 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5055 :
5056 : /* sort the list of SMgrRelations if necessary */
5057 26 : if (use_bsearch)
5058 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5059 :
5060 426010 : for (i = 0; i < NBuffers; i++)
5061 : {
5062 425984 : SMgrSortArray *srelent = NULL;
5063 425984 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5064 : uint32 buf_state;
5065 :
5066 : /*
5067 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5068 : * saves some cycles.
5069 : */
5070 :
5071 425984 : if (!use_bsearch)
5072 : {
5073 : int j;
5074 :
5075 875900 : for (j = 0; j < nrels; j++)
5076 : {
5077 458750 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5078 : {
5079 8834 : srelent = &srels[j];
5080 8834 : break;
5081 : }
5082 : }
5083 : }
5084 : else
5085 : {
5086 : RelFileLocator rlocator;
5087 :
5088 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5089 0 : srelent = bsearch(&rlocator,
5090 : srels, nrels, sizeof(SMgrSortArray),
5091 : rlocator_comparator);
5092 : }
5093 :
5094 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5095 425984 : if (srelent == NULL)
5096 417150 : continue;
5097 :
5098 : /* Make sure we can handle the pin */
5099 8834 : ReservePrivateRefCountEntry();
5100 8834 : ResourceOwnerEnlarge(CurrentResourceOwner);
5101 :
5102 8834 : buf_state = LockBufHdr(bufHdr);
5103 8834 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5104 8834 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5105 : {
5106 7918 : PinBuffer_Locked(bufHdr);
5107 7918 : FlushUnlockedBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5108 7918 : UnpinBuffer(bufHdr);
5109 : }
5110 : else
5111 916 : UnlockBufHdr(bufHdr);
5112 : }
5113 :
5114 26 : pfree(srels);
5115 : }
5116 :
5117 : /* ---------------------------------------------------------------------
5118 : * RelationCopyStorageUsingBuffer
5119 : *
5120 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5121 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5122 : *
5123 : * Refer comments atop CreateAndCopyRelationData() for details about
5124 : * 'permanent' parameter.
5125 : * --------------------------------------------------------------------
5126 : */
5127 : static void
5128 148452 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5129 : RelFileLocator dstlocator,
5130 : ForkNumber forkNum, bool permanent)
5131 : {
5132 : Buffer srcBuf;
5133 : Buffer dstBuf;
5134 : Page srcPage;
5135 : Page dstPage;
5136 : bool use_wal;
5137 : BlockNumber nblocks;
5138 : BlockNumber blkno;
5139 : PGIOAlignedBlock buf;
5140 : BufferAccessStrategy bstrategy_src;
5141 : BufferAccessStrategy bstrategy_dst;
5142 : BlockRangeReadStreamPrivate p;
5143 : ReadStream *src_stream;
5144 : SMgrRelation src_smgr;
5145 :
5146 : /*
5147 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5148 : * can skip it when copying any fork of an unlogged relation other than
5149 : * the init fork.
5150 : */
5151 148452 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5152 :
5153 : /* Get number of blocks in the source relation. */
5154 148452 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5155 : forkNum);
5156 :
5157 : /* Nothing to copy; just return. */
5158 148452 : if (nblocks == 0)
5159 25894 : return;
5160 :
5161 : /*
5162 : * Bulk extend the destination relation of the same size as the source
5163 : * relation before starting to copy block by block.
5164 : */
5165 122558 : memset(buf.data, 0, BLCKSZ);
5166 122558 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5167 : buf.data, true);
5168 :
5169 : /* This is a bulk operation, so use buffer access strategies. */
5170 122558 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5171 122558 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5172 :
5173 : /* Initialize streaming read */
5174 122558 : p.current_blocknum = 0;
5175 122558 : p.last_exclusive = nblocks;
5176 122558 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5177 :
5178 : /*
5179 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5180 : * locks.
5181 : */
5182 122558 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5183 : READ_STREAM_USE_BATCHING,
5184 : bstrategy_src,
5185 : src_smgr,
5186 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5187 : forkNum,
5188 : block_range_read_stream_cb,
5189 : &p,
5190 : 0);
5191 :
5192 : /* Iterate over each block of the source relation file. */
5193 590916 : for (blkno = 0; blkno < nblocks; blkno++)
5194 : {
5195 468362 : CHECK_FOR_INTERRUPTS();
5196 :
5197 : /* Read block from source relation. */
5198 468362 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5199 468358 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5200 468358 : srcPage = BufferGetPage(srcBuf);
5201 :
5202 468358 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5203 : BufferGetBlockNumber(srcBuf),
5204 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5205 : permanent);
5206 468358 : dstPage = BufferGetPage(dstBuf);
5207 :
5208 468358 : START_CRIT_SECTION();
5209 :
5210 : /* Copy page data from the source to the destination. */
5211 468358 : memcpy(dstPage, srcPage, BLCKSZ);
5212 468358 : MarkBufferDirty(dstBuf);
5213 :
5214 : /* WAL-log the copied page. */
5215 468358 : if (use_wal)
5216 259620 : log_newpage_buffer(dstBuf, true);
5217 :
5218 468358 : END_CRIT_SECTION();
5219 :
5220 468358 : UnlockReleaseBuffer(dstBuf);
5221 468358 : UnlockReleaseBuffer(srcBuf);
5222 : }
5223 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5224 122554 : read_stream_end(src_stream);
5225 :
5226 122554 : FreeAccessStrategy(bstrategy_src);
5227 122554 : FreeAccessStrategy(bstrategy_dst);
5228 : }
5229 :
5230 : /* ---------------------------------------------------------------------
5231 : * CreateAndCopyRelationData
5232 : *
5233 : * Create destination relation storage and copy all forks from the
5234 : * source relation to the destination.
5235 : *
5236 : * Pass permanent as true for permanent relations and false for
5237 : * unlogged relations. Currently this API is not supported for
5238 : * temporary relations.
5239 : * --------------------------------------------------------------------
5240 : */
5241 : void
5242 111592 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5243 : RelFileLocator dst_rlocator, bool permanent)
5244 : {
5245 : char relpersistence;
5246 : SMgrRelation src_rel;
5247 : SMgrRelation dst_rel;
5248 :
5249 : /* Set the relpersistence. */
5250 111592 : relpersistence = permanent ?
5251 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5252 :
5253 111592 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5254 111592 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5255 :
5256 : /*
5257 : * Create and copy all forks of the relation. During create database we
5258 : * have a separate cleanup mechanism which deletes complete database
5259 : * directory. Therefore, each individual relation doesn't need to be
5260 : * registered for cleanup.
5261 : */
5262 111592 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5263 :
5264 : /* copy main fork. */
5265 111592 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5266 : permanent);
5267 :
5268 : /* copy those extra forks that exist */
5269 111588 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5270 446352 : forkNum <= MAX_FORKNUM; forkNum++)
5271 : {
5272 334764 : if (smgrexists(src_rel, forkNum))
5273 : {
5274 36860 : smgrcreate(dst_rel, forkNum, false);
5275 :
5276 : /*
5277 : * WAL log creation if the relation is persistent, or this is the
5278 : * init fork of an unlogged relation.
5279 : */
5280 36860 : if (permanent || forkNum == INIT_FORKNUM)
5281 36860 : log_smgrcreate(&dst_rlocator, forkNum);
5282 :
5283 : /* Copy a fork's data, block by block. */
5284 36860 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5285 : permanent);
5286 : }
5287 : }
5288 111588 : }
5289 :
5290 : /* ---------------------------------------------------------------------
5291 : * FlushDatabaseBuffers
5292 : *
5293 : * This function writes all dirty pages of a database out to disk
5294 : * (or more accurately, out to kernel disk buffers), ensuring that the
5295 : * kernel has an up-to-date view of the database.
5296 : *
5297 : * Generally, the caller should be holding an appropriate lock to ensure
5298 : * no other backend is active in the target database; otherwise more
5299 : * pages could get dirtied.
5300 : *
5301 : * Note we don't worry about flushing any pages of temporary relations.
5302 : * It's assumed these wouldn't be interesting.
5303 : * --------------------------------------------------------------------
5304 : */
5305 : void
5306 10 : FlushDatabaseBuffers(Oid dbid)
5307 : {
5308 : int i;
5309 : BufferDesc *bufHdr;
5310 :
5311 1290 : for (i = 0; i < NBuffers; i++)
5312 : {
5313 : uint32 buf_state;
5314 :
5315 1280 : bufHdr = GetBufferDescriptor(i);
5316 :
5317 : /*
5318 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5319 : * saves some cycles.
5320 : */
5321 1280 : if (bufHdr->tag.dbOid != dbid)
5322 942 : continue;
5323 :
5324 : /* Make sure we can handle the pin */
5325 338 : ReservePrivateRefCountEntry();
5326 338 : ResourceOwnerEnlarge(CurrentResourceOwner);
5327 :
5328 338 : buf_state = LockBufHdr(bufHdr);
5329 338 : if (bufHdr->tag.dbOid == dbid &&
5330 338 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5331 : {
5332 32 : PinBuffer_Locked(bufHdr);
5333 32 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5334 32 : UnpinBuffer(bufHdr);
5335 : }
5336 : else
5337 306 : UnlockBufHdr(bufHdr);
5338 : }
5339 10 : }
5340 :
5341 : /*
5342 : * Flush a previously, shared or exclusively, locked and pinned buffer to the
5343 : * OS.
5344 : */
5345 : void
5346 160 : FlushOneBuffer(Buffer buffer)
5347 : {
5348 : BufferDesc *bufHdr;
5349 :
5350 : /* currently not needed, but no fundamental reason not to support */
5351 : Assert(!BufferIsLocal(buffer));
5352 :
5353 : Assert(BufferIsPinned(buffer));
5354 :
5355 160 : bufHdr = GetBufferDescriptor(buffer - 1);
5356 :
5357 : Assert(BufferIsLockedByMe(buffer));
5358 :
5359 160 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5360 160 : }
5361 :
5362 : /*
5363 : * ReleaseBuffer -- release the pin on a buffer
5364 : */
5365 : void
5366 125442454 : ReleaseBuffer(Buffer buffer)
5367 : {
5368 125442454 : if (!BufferIsValid(buffer))
5369 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5370 :
5371 125442454 : if (BufferIsLocal(buffer))
5372 3220342 : UnpinLocalBuffer(buffer);
5373 : else
5374 122222112 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5375 125442454 : }
5376 :
5377 : /*
5378 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5379 : *
5380 : * This is just a shorthand for a common combination.
5381 : */
5382 : void
5383 37950326 : UnlockReleaseBuffer(Buffer buffer)
5384 : {
5385 37950326 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5386 37950326 : ReleaseBuffer(buffer);
5387 37950326 : }
5388 :
5389 : /*
5390 : * IncrBufferRefCount
5391 : * Increment the pin count on a buffer that we have *already* pinned
5392 : * at least once.
5393 : *
5394 : * This function cannot be used on a buffer we do not have pinned,
5395 : * because it doesn't change the shared buffer state.
5396 : */
5397 : void
5398 23369944 : IncrBufferRefCount(Buffer buffer)
5399 : {
5400 : Assert(BufferIsPinned(buffer));
5401 23369944 : ResourceOwnerEnlarge(CurrentResourceOwner);
5402 23369944 : if (BufferIsLocal(buffer))
5403 709372 : LocalRefCount[-buffer - 1]++;
5404 : else
5405 : {
5406 : PrivateRefCountEntry *ref;
5407 :
5408 22660572 : ref = GetPrivateRefCountEntry(buffer, true);
5409 : Assert(ref != NULL);
5410 22660572 : ref->refcount++;
5411 : }
5412 23369944 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5413 23369944 : }
5414 :
5415 : /*
5416 : * MarkBufferDirtyHint
5417 : *
5418 : * Mark a buffer dirty for non-critical changes.
5419 : *
5420 : * This is essentially the same as MarkBufferDirty, except:
5421 : *
5422 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5423 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5424 : * 2. The caller might have only share-lock instead of exclusive-lock on the
5425 : * buffer's content lock.
5426 : * 3. This function does not guarantee that the buffer is always marked dirty
5427 : * (due to a race condition), so it cannot be used for important changes.
5428 : */
5429 : void
5430 19966062 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5431 : {
5432 : BufferDesc *bufHdr;
5433 19966062 : Page page = BufferGetPage(buffer);
5434 :
5435 19966062 : if (!BufferIsValid(buffer))
5436 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5437 :
5438 19966062 : if (BufferIsLocal(buffer))
5439 : {
5440 1270290 : MarkLocalBufferDirty(buffer);
5441 1270290 : return;
5442 : }
5443 :
5444 18695772 : bufHdr = GetBufferDescriptor(buffer - 1);
5445 :
5446 : Assert(GetPrivateRefCount(buffer) > 0);
5447 : /* here, either share or exclusive lock is OK */
5448 : Assert(BufferIsLockedByMe(buffer));
5449 :
5450 : /*
5451 : * This routine might get called many times on the same page, if we are
5452 : * making the first scan after commit of an xact that added/deleted many
5453 : * tuples. So, be as quick as we can if the buffer is already dirty. We
5454 : * do this by not acquiring spinlock if it looks like the status bits are
5455 : * already set. Since we make this test unlocked, there's a chance we
5456 : * might fail to notice that the flags have just been cleared, and failed
5457 : * to reset them, due to memory-ordering issues. But since this function
5458 : * is only intended to be used in cases where failing to write out the
5459 : * data would be harmless anyway, it doesn't really matter.
5460 : */
5461 18695772 : if ((pg_atomic_read_u32(&bufHdr->state) & (BM_DIRTY | BM_JUST_DIRTIED)) !=
5462 : (BM_DIRTY | BM_JUST_DIRTIED))
5463 : {
5464 1752910 : XLogRecPtr lsn = InvalidXLogRecPtr;
5465 1752910 : bool dirtied = false;
5466 1752910 : bool delayChkptFlags = false;
5467 : uint32 buf_state;
5468 :
5469 : /*
5470 : * If we need to protect hint bit updates from torn writes, WAL-log a
5471 : * full page image of the page. This full page image is only necessary
5472 : * if the hint bit update is the first change to the page since the
5473 : * last checkpoint.
5474 : *
5475 : * We don't check full_page_writes here because that logic is included
5476 : * when we call XLogInsert() since the value changes dynamically.
5477 : */
5478 3503630 : if (XLogHintBitIsNeeded() &&
5479 1750720 : (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT))
5480 : {
5481 : /*
5482 : * If we must not write WAL, due to a relfilelocator-specific
5483 : * condition or being in recovery, don't dirty the page. We can
5484 : * set the hint, just not dirty the page as a result so the hint
5485 : * is lost when we evict the page or shutdown.
5486 : *
5487 : * See src/backend/storage/page/README for longer discussion.
5488 : */
5489 1874640 : if (RecoveryInProgress() ||
5490 123984 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5491 1632366 : return;
5492 :
5493 : /*
5494 : * If the block is already dirty because we either made a change
5495 : * or set a hint already, then we don't need to write a full page
5496 : * image. Note that aggressive cleaning of blocks dirtied by hint
5497 : * bit setting would increase the call rate. Bulk setting of hint
5498 : * bits would reduce the call rate...
5499 : *
5500 : * We must issue the WAL record before we mark the buffer dirty.
5501 : * Otherwise we might write the page before we write the WAL. That
5502 : * causes a race condition, since a checkpoint might occur between
5503 : * writing the WAL record and marking the buffer dirty. We solve
5504 : * that with a kluge, but one that is already in use during
5505 : * transaction commit to prevent race conditions. Basically, we
5506 : * simply prevent the checkpoint WAL record from being written
5507 : * until we have marked the buffer dirty. We don't start the
5508 : * checkpoint flush until we have marked dirty, so our checkpoint
5509 : * must flush the change to disk successfully or the checkpoint
5510 : * never gets written, so crash recovery will fix.
5511 : *
5512 : * It's possible we may enter here without an xid, so it is
5513 : * essential that CreateCheckPoint waits for virtual transactions
5514 : * rather than full transactionids.
5515 : */
5516 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
5517 118290 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
5518 118290 : delayChkptFlags = true;
5519 118290 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5520 : }
5521 :
5522 120544 : buf_state = LockBufHdr(bufHdr);
5523 :
5524 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5525 :
5526 120544 : if (!(buf_state & BM_DIRTY))
5527 : {
5528 120468 : dirtied = true; /* Means "will be dirtied by this action" */
5529 :
5530 : /*
5531 : * Set the page LSN if we wrote a backup block. We aren't supposed
5532 : * to set this when only holding a share lock but as long as we
5533 : * serialise it somehow we're OK. We choose to set LSN while
5534 : * holding the buffer header lock, which causes any reader of an
5535 : * LSN who holds only a share lock to also obtain a buffer header
5536 : * lock before using PageGetLSN(), which is enforced in
5537 : * BufferGetLSNAtomic().
5538 : *
5539 : * If checksums are enabled, you might think we should reset the
5540 : * checksum here. That will happen when the page is written
5541 : * sometime later in this checkpoint cycle.
5542 : */
5543 120468 : if (XLogRecPtrIsValid(lsn))
5544 62584 : PageSetLSN(page, lsn);
5545 : }
5546 :
5547 120544 : UnlockBufHdrExt(bufHdr, buf_state,
5548 : BM_DIRTY | BM_JUST_DIRTIED,
5549 : 0, 0);
5550 :
5551 120544 : if (delayChkptFlags)
5552 118290 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
5553 :
5554 120544 : if (dirtied)
5555 : {
5556 120468 : pgBufferUsage.shared_blks_dirtied++;
5557 120468 : if (VacuumCostActive)
5558 2762 : VacuumCostBalance += VacuumCostPageDirty;
5559 : }
5560 : }
5561 : }
5562 :
5563 : /*
5564 : * Release buffer content locks for shared buffers.
5565 : *
5566 : * Used to clean up after errors.
5567 : *
5568 : * Currently, we can expect that lwlock.c's LWLockReleaseAll() took care
5569 : * of releasing buffer content locks per se; the only thing we need to deal
5570 : * with here is clearing any PIN_COUNT request that was in progress.
5571 : */
5572 : void
5573 104938 : UnlockBuffers(void)
5574 : {
5575 104938 : BufferDesc *buf = PinCountWaitBuf;
5576 :
5577 104938 : if (buf)
5578 : {
5579 : uint32 buf_state;
5580 0 : uint32 unset_bits = 0;
5581 :
5582 0 : buf_state = LockBufHdr(buf);
5583 :
5584 : /*
5585 : * Don't complain if flag bit not set; it could have been reset but we
5586 : * got a cancel/die interrupt before getting the signal.
5587 : */
5588 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5589 0 : buf->wait_backend_pgprocno == MyProcNumber)
5590 0 : unset_bits = BM_PIN_COUNT_WAITER;
5591 :
5592 0 : UnlockBufHdrExt(buf, buf_state,
5593 : 0, unset_bits,
5594 : 0);
5595 :
5596 0 : PinCountWaitBuf = NULL;
5597 : }
5598 104938 : }
5599 :
5600 : /*
5601 : * Acquire or release the content_lock for the buffer.
5602 : */
5603 : void
5604 346855492 : LockBuffer(Buffer buffer, int mode)
5605 : {
5606 : BufferDesc *buf;
5607 :
5608 : Assert(BufferIsPinned(buffer));
5609 346855492 : if (BufferIsLocal(buffer))
5610 19800312 : return; /* local buffers need no lock */
5611 :
5612 327055180 : buf = GetBufferDescriptor(buffer - 1);
5613 :
5614 327055180 : if (mode == BUFFER_LOCK_UNLOCK)
5615 165001860 : LWLockRelease(BufferDescriptorGetContentLock(buf));
5616 162053320 : else if (mode == BUFFER_LOCK_SHARE)
5617 114665262 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
5618 47388058 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
5619 47388058 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_EXCLUSIVE);
5620 : else
5621 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
5622 : }
5623 :
5624 : /*
5625 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
5626 : *
5627 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
5628 : */
5629 : bool
5630 3081990 : ConditionalLockBuffer(Buffer buffer)
5631 : {
5632 : BufferDesc *buf;
5633 :
5634 : Assert(BufferIsPinned(buffer));
5635 3081990 : if (BufferIsLocal(buffer))
5636 129320 : return true; /* act as though we got it */
5637 :
5638 2952670 : buf = GetBufferDescriptor(buffer - 1);
5639 :
5640 2952670 : return LWLockConditionalAcquire(BufferDescriptorGetContentLock(buf),
5641 : LW_EXCLUSIVE);
5642 : }
5643 :
5644 : /*
5645 : * Verify that this backend is pinning the buffer exactly once.
5646 : *
5647 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
5648 : * holds a pin on the buffer. We do not care whether some other backend does.
5649 : */
5650 : void
5651 5058194 : CheckBufferIsPinnedOnce(Buffer buffer)
5652 : {
5653 5058194 : if (BufferIsLocal(buffer))
5654 : {
5655 1578 : if (LocalRefCount[-buffer - 1] != 1)
5656 0 : elog(ERROR, "incorrect local pin count: %d",
5657 : LocalRefCount[-buffer - 1]);
5658 : }
5659 : else
5660 : {
5661 5056616 : if (GetPrivateRefCount(buffer) != 1)
5662 0 : elog(ERROR, "incorrect local pin count: %d",
5663 : GetPrivateRefCount(buffer));
5664 : }
5665 5058194 : }
5666 :
5667 : /*
5668 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
5669 : *
5670 : * Items may be deleted from a disk page only when the caller (a) holds an
5671 : * exclusive lock on the buffer and (b) has observed that no other backend
5672 : * holds a pin on the buffer. If there is a pin, then the other backend
5673 : * might have a pointer into the buffer (for example, a heapscan reference
5674 : * to an item --- see README for more details). It's OK if a pin is added
5675 : * after the cleanup starts, however; the newly-arrived backend will be
5676 : * unable to look at the page until we release the exclusive lock.
5677 : *
5678 : * To implement this protocol, a would-be deleter must pin the buffer and
5679 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
5680 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
5681 : * it has successfully observed pin count = 1.
5682 : */
5683 : void
5684 44788 : LockBufferForCleanup(Buffer buffer)
5685 : {
5686 : BufferDesc *bufHdr;
5687 44788 : TimestampTz waitStart = 0;
5688 44788 : bool waiting = false;
5689 44788 : bool logged_recovery_conflict = false;
5690 :
5691 : Assert(BufferIsPinned(buffer));
5692 : Assert(PinCountWaitBuf == NULL);
5693 :
5694 44788 : CheckBufferIsPinnedOnce(buffer);
5695 :
5696 : /*
5697 : * We do not yet need to be worried about in-progress AIOs holding a pin,
5698 : * as we, so far, only support doing reads via AIO and this function can
5699 : * only be called once the buffer is valid (i.e. no read can be in
5700 : * flight).
5701 : */
5702 :
5703 : /* Nobody else to wait for */
5704 44788 : if (BufferIsLocal(buffer))
5705 32 : return;
5706 :
5707 44756 : bufHdr = GetBufferDescriptor(buffer - 1);
5708 :
5709 : for (;;)
5710 34 : {
5711 : uint32 buf_state;
5712 44790 : uint32 unset_bits = 0;
5713 :
5714 : /* Try to acquire lock */
5715 44790 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5716 44790 : buf_state = LockBufHdr(bufHdr);
5717 :
5718 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5719 44790 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5720 : {
5721 : /* Successfully acquired exclusive lock with pincount 1 */
5722 44756 : UnlockBufHdr(bufHdr);
5723 :
5724 : /*
5725 : * Emit the log message if recovery conflict on buffer pin was
5726 : * resolved but the startup process waited longer than
5727 : * deadlock_timeout for it.
5728 : */
5729 44756 : if (logged_recovery_conflict)
5730 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5731 : waitStart, GetCurrentTimestamp(),
5732 : NULL, false);
5733 :
5734 44756 : if (waiting)
5735 : {
5736 : /* reset ps display to remove the suffix if we added one */
5737 4 : set_ps_display_remove_suffix();
5738 4 : waiting = false;
5739 : }
5740 44756 : return;
5741 : }
5742 : /* Failed, so mark myself as waiting for pincount 1 */
5743 34 : if (buf_state & BM_PIN_COUNT_WAITER)
5744 : {
5745 0 : UnlockBufHdr(bufHdr);
5746 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5747 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
5748 : }
5749 34 : bufHdr->wait_backend_pgprocno = MyProcNumber;
5750 34 : PinCountWaitBuf = bufHdr;
5751 34 : UnlockBufHdrExt(bufHdr, buf_state,
5752 : BM_PIN_COUNT_WAITER, 0,
5753 : 0);
5754 34 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5755 :
5756 : /* Wait to be signaled by UnpinBuffer() */
5757 34 : if (InHotStandby)
5758 : {
5759 22 : if (!waiting)
5760 : {
5761 : /* adjust the process title to indicate that it's waiting */
5762 4 : set_ps_display_suffix("waiting");
5763 4 : waiting = true;
5764 : }
5765 :
5766 : /*
5767 : * Emit the log message if the startup process is waiting longer
5768 : * than deadlock_timeout for recovery conflict on buffer pin.
5769 : *
5770 : * Skip this if first time through because the startup process has
5771 : * not started waiting yet in this case. So, the wait start
5772 : * timestamp is set after this logic.
5773 : */
5774 22 : if (waitStart != 0 && !logged_recovery_conflict)
5775 : {
5776 8 : TimestampTz now = GetCurrentTimestamp();
5777 :
5778 8 : if (TimestampDifferenceExceeds(waitStart, now,
5779 : DeadlockTimeout))
5780 : {
5781 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5782 : waitStart, now, NULL, true);
5783 4 : logged_recovery_conflict = true;
5784 : }
5785 : }
5786 :
5787 : /*
5788 : * Set the wait start timestamp if logging is enabled and first
5789 : * time through.
5790 : */
5791 22 : if (log_recovery_conflict_waits && waitStart == 0)
5792 4 : waitStart = GetCurrentTimestamp();
5793 :
5794 : /* Publish the bufid that Startup process waits on */
5795 22 : SetStartupBufferPinWaitBufId(buffer - 1);
5796 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
5797 22 : ResolveRecoveryConflictWithBufferPin();
5798 : /* Reset the published bufid */
5799 22 : SetStartupBufferPinWaitBufId(-1);
5800 : }
5801 : else
5802 12 : ProcWaitForSignal(WAIT_EVENT_BUFFER_PIN);
5803 :
5804 : /*
5805 : * Remove flag marking us as waiter. Normally this will not be set
5806 : * anymore, but ProcWaitForSignal() can return for other signals as
5807 : * well. We take care to only reset the flag if we're the waiter, as
5808 : * theoretically another backend could have started waiting. That's
5809 : * impossible with the current usages due to table level locking, but
5810 : * better be safe.
5811 : */
5812 34 : buf_state = LockBufHdr(bufHdr);
5813 34 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5814 20 : bufHdr->wait_backend_pgprocno == MyProcNumber)
5815 20 : unset_bits |= BM_PIN_COUNT_WAITER;
5816 :
5817 34 : UnlockBufHdrExt(bufHdr, buf_state,
5818 : 0, unset_bits,
5819 : 0);
5820 :
5821 34 : PinCountWaitBuf = NULL;
5822 : /* Loop back and try again */
5823 : }
5824 : }
5825 :
5826 : /*
5827 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
5828 : * requests cancellation of all pin holders that are blocking it.
5829 : */
5830 : bool
5831 8 : HoldingBufferPinThatDelaysRecovery(void)
5832 : {
5833 8 : int bufid = GetStartupBufferPinWaitBufId();
5834 :
5835 : /*
5836 : * If we get woken slowly then it's possible that the Startup process was
5837 : * already woken by other backends before we got here. Also possible that
5838 : * we get here by multiple interrupts or interrupts at inappropriate
5839 : * times, so make sure we do nothing if the bufid is not set.
5840 : */
5841 8 : if (bufid < 0)
5842 4 : return false;
5843 :
5844 4 : if (GetPrivateRefCount(bufid + 1) > 0)
5845 4 : return true;
5846 :
5847 0 : return false;
5848 : }
5849 :
5850 : /*
5851 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
5852 : *
5853 : * We won't loop, but just check once to see if the pin count is OK. If
5854 : * not, return false with no lock held.
5855 : */
5856 : bool
5857 1245150 : ConditionalLockBufferForCleanup(Buffer buffer)
5858 : {
5859 : BufferDesc *bufHdr;
5860 : uint32 buf_state,
5861 : refcount;
5862 :
5863 : Assert(BufferIsValid(buffer));
5864 :
5865 : /* see AIO related comment in LockBufferForCleanup() */
5866 :
5867 1245150 : if (BufferIsLocal(buffer))
5868 : {
5869 1608 : refcount = LocalRefCount[-buffer - 1];
5870 : /* There should be exactly one pin */
5871 : Assert(refcount > 0);
5872 1608 : if (refcount != 1)
5873 42 : return false;
5874 : /* Nobody else to wait for */
5875 1566 : return true;
5876 : }
5877 :
5878 : /* There should be exactly one local pin */
5879 1243542 : refcount = GetPrivateRefCount(buffer);
5880 : Assert(refcount);
5881 1243542 : if (refcount != 1)
5882 532 : return false;
5883 :
5884 : /* Try to acquire lock */
5885 1243010 : if (!ConditionalLockBuffer(buffer))
5886 186 : return false;
5887 :
5888 1242824 : bufHdr = GetBufferDescriptor(buffer - 1);
5889 1242824 : buf_state = LockBufHdr(bufHdr);
5890 1242824 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
5891 :
5892 : Assert(refcount > 0);
5893 1242824 : if (refcount == 1)
5894 : {
5895 : /* Successfully acquired exclusive lock with pincount 1 */
5896 1242182 : UnlockBufHdr(bufHdr);
5897 1242182 : return true;
5898 : }
5899 :
5900 : /* Failed, so release the lock */
5901 642 : UnlockBufHdr(bufHdr);
5902 642 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5903 642 : return false;
5904 : }
5905 :
5906 : /*
5907 : * IsBufferCleanupOK - as above, but we already have the lock
5908 : *
5909 : * Check whether it's OK to perform cleanup on a buffer we've already
5910 : * locked. If we observe that the pin count is 1, our exclusive lock
5911 : * happens to be a cleanup lock, and we can proceed with anything that
5912 : * would have been allowable had we sought a cleanup lock originally.
5913 : */
5914 : bool
5915 3324 : IsBufferCleanupOK(Buffer buffer)
5916 : {
5917 : BufferDesc *bufHdr;
5918 : uint32 buf_state;
5919 :
5920 : Assert(BufferIsValid(buffer));
5921 :
5922 : /* see AIO related comment in LockBufferForCleanup() */
5923 :
5924 3324 : if (BufferIsLocal(buffer))
5925 : {
5926 : /* There should be exactly one pin */
5927 0 : if (LocalRefCount[-buffer - 1] != 1)
5928 0 : return false;
5929 : /* Nobody else to wait for */
5930 0 : return true;
5931 : }
5932 :
5933 : /* There should be exactly one local pin */
5934 3324 : if (GetPrivateRefCount(buffer) != 1)
5935 0 : return false;
5936 :
5937 3324 : bufHdr = GetBufferDescriptor(buffer - 1);
5938 :
5939 : /* caller must hold exclusive lock on buffer */
5940 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
5941 :
5942 3324 : buf_state = LockBufHdr(bufHdr);
5943 :
5944 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5945 3324 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5946 : {
5947 : /* pincount is OK. */
5948 3324 : UnlockBufHdr(bufHdr);
5949 3324 : return true;
5950 : }
5951 :
5952 0 : UnlockBufHdr(bufHdr);
5953 0 : return false;
5954 : }
5955 :
5956 :
5957 : /*
5958 : * Functions for buffer I/O handling
5959 : *
5960 : * Also note that these are used only for shared buffers, not local ones.
5961 : */
5962 :
5963 : /*
5964 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
5965 : */
5966 : static void
5967 138144 : WaitIO(BufferDesc *buf)
5968 : {
5969 138144 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
5970 :
5971 138144 : ConditionVariablePrepareToSleep(cv);
5972 : for (;;)
5973 14014 : {
5974 : uint32 buf_state;
5975 : PgAioWaitRef iow;
5976 :
5977 : /*
5978 : * It may not be necessary to acquire the spinlock to check the flag
5979 : * here, but since this test is essential for correctness, we'd better
5980 : * play it safe.
5981 : */
5982 152158 : buf_state = LockBufHdr(buf);
5983 :
5984 : /*
5985 : * Copy the wait reference while holding the spinlock. This protects
5986 : * against a concurrent TerminateBufferIO() in another backend from
5987 : * clearing the wref while it's being read.
5988 : */
5989 152158 : iow = buf->io_wref;
5990 152158 : UnlockBufHdr(buf);
5991 :
5992 : /* no IO in progress, we don't need to wait */
5993 152158 : if (!(buf_state & BM_IO_IN_PROGRESS))
5994 138144 : break;
5995 :
5996 : /*
5997 : * The buffer has asynchronous IO in progress, wait for it to
5998 : * complete.
5999 : */
6000 14014 : if (pgaio_wref_valid(&iow))
6001 : {
6002 11338 : pgaio_wref_wait(&iow);
6003 :
6004 : /*
6005 : * The AIO subsystem internally uses condition variables and thus
6006 : * might remove this backend from the BufferDesc's CV. While that
6007 : * wouldn't cause a correctness issue (the first CV sleep just
6008 : * immediately returns if not already registered), it seems worth
6009 : * avoiding unnecessary loop iterations, given that we take care
6010 : * to do so at the start of the function.
6011 : */
6012 11338 : ConditionVariablePrepareToSleep(cv);
6013 11338 : continue;
6014 : }
6015 :
6016 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
6017 2676 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
6018 : }
6019 138144 : ConditionVariableCancelSleep();
6020 138144 : }
6021 :
6022 : /*
6023 : * StartBufferIO: begin I/O on this buffer
6024 : * (Assumptions)
6025 : * My process is executing no IO on this buffer
6026 : * The buffer is Pinned
6027 : *
6028 : * In some scenarios multiple backends could attempt the same I/O operation
6029 : * concurrently. If someone else has already started I/O on this buffer then
6030 : * we will wait for completion of the IO using WaitIO().
6031 : *
6032 : * Input operations are only attempted on buffers that are not BM_VALID,
6033 : * and output operations only on buffers that are BM_VALID and BM_DIRTY,
6034 : * so we can always tell if the work is already done.
6035 : *
6036 : * Returns true if we successfully marked the buffer as I/O busy,
6037 : * false if someone else already did the work.
6038 : *
6039 : * If nowait is true, then we don't wait for an I/O to be finished by another
6040 : * backend. In that case, false indicates either that the I/O was already
6041 : * finished, or is still in progress. This is useful for callers that want to
6042 : * find out if they can perform the I/O as part of a larger operation, without
6043 : * waiting for the answer or distinguishing the reasons why not.
6044 : */
6045 : bool
6046 4975718 : StartBufferIO(BufferDesc *buf, bool forInput, bool nowait)
6047 : {
6048 : uint32 buf_state;
6049 :
6050 4975718 : ResourceOwnerEnlarge(CurrentResourceOwner);
6051 :
6052 : for (;;)
6053 : {
6054 4989814 : buf_state = LockBufHdr(buf);
6055 :
6056 4989814 : if (!(buf_state & BM_IO_IN_PROGRESS))
6057 4975710 : break;
6058 14104 : UnlockBufHdr(buf);
6059 14104 : if (nowait)
6060 8 : return false;
6061 14096 : WaitIO(buf);
6062 : }
6063 :
6064 : /* Once we get here, there is definitely no I/O active on this buffer */
6065 :
6066 : /* Check if someone else already did the I/O */
6067 4975710 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
6068 : {
6069 15002 : UnlockBufHdr(buf);
6070 15002 : return false;
6071 : }
6072 :
6073 4960708 : UnlockBufHdrExt(buf, buf_state,
6074 : BM_IO_IN_PROGRESS, 0,
6075 : 0);
6076 :
6077 4960708 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
6078 : BufferDescriptorGetBuffer(buf));
6079 :
6080 4960708 : return true;
6081 : }
6082 :
6083 : /*
6084 : * TerminateBufferIO: release a buffer we were doing I/O on
6085 : * (Assumptions)
6086 : * My process is executing IO for the buffer
6087 : * BM_IO_IN_PROGRESS bit is set for the buffer
6088 : * The buffer is Pinned
6089 : *
6090 : * If clear_dirty is true and BM_JUST_DIRTIED is not set, we clear the
6091 : * buffer's BM_DIRTY flag. This is appropriate when terminating a
6092 : * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
6093 : * marking the buffer clean if it was re-dirtied while we were writing.
6094 : *
6095 : * set_flag_bits gets ORed into the buffer's flags. It must include
6096 : * BM_IO_ERROR in a failure case. For successful completion it could
6097 : * be 0, or BM_VALID if we just finished reading in the page.
6098 : *
6099 : * If forget_owner is true, we release the buffer I/O from the current
6100 : * resource owner. (forget_owner=false is used when the resource owner itself
6101 : * is being released)
6102 : */
6103 : void
6104 4685246 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint32 set_flag_bits,
6105 : bool forget_owner, bool release_aio)
6106 : {
6107 : uint32 buf_state;
6108 4685246 : uint32 unset_flag_bits = 0;
6109 4685246 : int refcount_change = 0;
6110 :
6111 4685246 : buf_state = LockBufHdr(buf);
6112 :
6113 : Assert(buf_state & BM_IO_IN_PROGRESS);
6114 4685246 : unset_flag_bits |= BM_IO_IN_PROGRESS;
6115 :
6116 : /* Clear earlier errors, if this IO failed, it'll be marked again */
6117 4685246 : unset_flag_bits |= BM_IO_ERROR;
6118 :
6119 4685246 : if (clear_dirty && !(buf_state & BM_JUST_DIRTIED))
6120 1122644 : unset_flag_bits |= BM_DIRTY | BM_CHECKPOINT_NEEDED;
6121 :
6122 4685246 : if (release_aio)
6123 : {
6124 : /* release ownership by the AIO subsystem */
6125 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6126 2555494 : refcount_change = -1;
6127 2555494 : pgaio_wref_clear(&buf->io_wref);
6128 : }
6129 :
6130 4685246 : buf_state = UnlockBufHdrExt(buf, buf_state,
6131 : set_flag_bits, unset_flag_bits,
6132 : refcount_change);
6133 :
6134 4685246 : if (forget_owner)
6135 2129710 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
6136 : BufferDescriptorGetBuffer(buf));
6137 :
6138 4685246 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
6139 :
6140 : /*
6141 : * Support LockBufferForCleanup()
6142 : *
6143 : * We may have just released the last pin other than the waiter's. In most
6144 : * cases, this backend holds another pin on the buffer. But, if, for
6145 : * example, this backend is completing an IO issued by another backend, it
6146 : * may be time to wake the waiter.
6147 : */
6148 4685246 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
6149 0 : WakePinCountWaiter(buf);
6150 4685246 : }
6151 :
6152 : /*
6153 : * AbortBufferIO: Clean up active buffer I/O after an error.
6154 : *
6155 : * All LWLocks we might have held have been released,
6156 : * but we haven't yet released buffer pins, so the buffer is still pinned.
6157 : *
6158 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
6159 : * possible the error condition wasn't related to the I/O.
6160 : *
6161 : * Note: this does not remove the buffer I/O from the resource owner.
6162 : * That's correct when we're releasing the whole resource owner, but
6163 : * beware if you use this in other contexts.
6164 : */
6165 : static void
6166 30 : AbortBufferIO(Buffer buffer)
6167 : {
6168 30 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
6169 : uint32 buf_state;
6170 :
6171 30 : buf_state = LockBufHdr(buf_hdr);
6172 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
6173 :
6174 30 : if (!(buf_state & BM_VALID))
6175 : {
6176 : Assert(!(buf_state & BM_DIRTY));
6177 30 : UnlockBufHdr(buf_hdr);
6178 : }
6179 : else
6180 : {
6181 : Assert(buf_state & BM_DIRTY);
6182 0 : UnlockBufHdr(buf_hdr);
6183 :
6184 : /* Issue notice if this is not the first failure... */
6185 0 : if (buf_state & BM_IO_ERROR)
6186 : {
6187 : /* Buffer is pinned, so we can read tag without spinlock */
6188 0 : ereport(WARNING,
6189 : (errcode(ERRCODE_IO_ERROR),
6190 : errmsg("could not write block %u of %s",
6191 : buf_hdr->tag.blockNum,
6192 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
6193 : BufTagGetForkNum(&buf_hdr->tag)).str),
6194 : errdetail("Multiple failures --- write error might be permanent.")));
6195 : }
6196 : }
6197 :
6198 30 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
6199 30 : }
6200 :
6201 : /*
6202 : * Error context callback for errors occurring during shared buffer writes.
6203 : */
6204 : static void
6205 68 : shared_buffer_write_error_callback(void *arg)
6206 : {
6207 68 : BufferDesc *bufHdr = (BufferDesc *) arg;
6208 :
6209 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
6210 68 : if (bufHdr != NULL)
6211 136 : errcontext("writing block %u of relation \"%s\"",
6212 : bufHdr->tag.blockNum,
6213 68 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
6214 : BufTagGetForkNum(&bufHdr->tag)).str);
6215 68 : }
6216 :
6217 : /*
6218 : * Error context callback for errors occurring during local buffer writes.
6219 : */
6220 : static void
6221 0 : local_buffer_write_error_callback(void *arg)
6222 : {
6223 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
6224 :
6225 0 : if (bufHdr != NULL)
6226 0 : errcontext("writing block %u of relation \"%s\"",
6227 : bufHdr->tag.blockNum,
6228 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
6229 : MyProcNumber,
6230 : BufTagGetForkNum(&bufHdr->tag)).str);
6231 0 : }
6232 :
6233 : /*
6234 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
6235 : */
6236 : static int
6237 19417632 : rlocator_comparator(const void *p1, const void *p2)
6238 : {
6239 19417632 : RelFileLocator n1 = *(const RelFileLocator *) p1;
6240 19417632 : RelFileLocator n2 = *(const RelFileLocator *) p2;
6241 :
6242 19417632 : if (n1.relNumber < n2.relNumber)
6243 19347296 : return -1;
6244 70336 : else if (n1.relNumber > n2.relNumber)
6245 66758 : return 1;
6246 :
6247 3578 : if (n1.dbOid < n2.dbOid)
6248 0 : return -1;
6249 3578 : else if (n1.dbOid > n2.dbOid)
6250 0 : return 1;
6251 :
6252 3578 : if (n1.spcOid < n2.spcOid)
6253 0 : return -1;
6254 3578 : else if (n1.spcOid > n2.spcOid)
6255 0 : return 1;
6256 : else
6257 3578 : return 0;
6258 : }
6259 :
6260 : /*
6261 : * Lock buffer header - set BM_LOCKED in buffer state.
6262 : */
6263 : uint32
6264 64307780 : LockBufHdr(BufferDesc *desc)
6265 : {
6266 : SpinDelayStatus delayStatus;
6267 : uint32 old_buf_state;
6268 :
6269 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
6270 :
6271 64307780 : init_local_spin_delay(&delayStatus);
6272 :
6273 : while (true)
6274 : {
6275 : /* set BM_LOCKED flag */
6276 64313078 : old_buf_state = pg_atomic_fetch_or_u32(&desc->state, BM_LOCKED);
6277 : /* if it wasn't set before we're OK */
6278 64313078 : if (!(old_buf_state & BM_LOCKED))
6279 64307780 : break;
6280 5298 : perform_spin_delay(&delayStatus);
6281 : }
6282 64307780 : finish_spin_delay(&delayStatus);
6283 64307780 : return old_buf_state | BM_LOCKED;
6284 : }
6285 :
6286 : /*
6287 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
6288 : * state at that point.
6289 : *
6290 : * Obviously the buffer could be locked by the time the value is returned, so
6291 : * this is primarily useful in CAS style loops.
6292 : */
6293 : pg_noinline uint32
6294 1692 : WaitBufHdrUnlocked(BufferDesc *buf)
6295 : {
6296 : SpinDelayStatus delayStatus;
6297 : uint32 buf_state;
6298 :
6299 1692 : init_local_spin_delay(&delayStatus);
6300 :
6301 1692 : buf_state = pg_atomic_read_u32(&buf->state);
6302 :
6303 40402 : while (buf_state & BM_LOCKED)
6304 : {
6305 38710 : perform_spin_delay(&delayStatus);
6306 38710 : buf_state = pg_atomic_read_u32(&buf->state);
6307 : }
6308 :
6309 1692 : finish_spin_delay(&delayStatus);
6310 :
6311 1692 : return buf_state;
6312 : }
6313 :
6314 : /*
6315 : * BufferTag comparator.
6316 : */
6317 : static inline int
6318 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
6319 : {
6320 : int ret;
6321 : RelFileLocator rlocatora;
6322 : RelFileLocator rlocatorb;
6323 :
6324 0 : rlocatora = BufTagGetRelFileLocator(ba);
6325 0 : rlocatorb = BufTagGetRelFileLocator(bb);
6326 :
6327 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
6328 :
6329 0 : if (ret != 0)
6330 0 : return ret;
6331 :
6332 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
6333 0 : return -1;
6334 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
6335 0 : return 1;
6336 :
6337 0 : if (ba->blockNum < bb->blockNum)
6338 0 : return -1;
6339 0 : if (ba->blockNum > bb->blockNum)
6340 0 : return 1;
6341 :
6342 0 : return 0;
6343 : }
6344 :
6345 : /*
6346 : * Comparator determining the writeout order in a checkpoint.
6347 : *
6348 : * It is important that tablespaces are compared first, the logic balancing
6349 : * writes between tablespaces relies on it.
6350 : */
6351 : static inline int
6352 5977248 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
6353 : {
6354 : /* compare tablespace */
6355 5977248 : if (a->tsId < b->tsId)
6356 13888 : return -1;
6357 5963360 : else if (a->tsId > b->tsId)
6358 52278 : return 1;
6359 : /* compare relation */
6360 5911082 : if (a->relNumber < b->relNumber)
6361 1665376 : return -1;
6362 4245706 : else if (a->relNumber > b->relNumber)
6363 1623500 : return 1;
6364 : /* compare fork */
6365 2622206 : else if (a->forkNum < b->forkNum)
6366 111586 : return -1;
6367 2510620 : else if (a->forkNum > b->forkNum)
6368 119382 : return 1;
6369 : /* compare block number */
6370 2391238 : else if (a->blockNum < b->blockNum)
6371 1173016 : return -1;
6372 1218222 : else if (a->blockNum > b->blockNum)
6373 1145100 : return 1;
6374 : /* equal page IDs are unlikely, but not impossible */
6375 73122 : return 0;
6376 : }
6377 :
6378 : /*
6379 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
6380 : * progress.
6381 : */
6382 : static int
6383 486888 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
6384 : {
6385 486888 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
6386 486888 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
6387 :
6388 : /* we want a min-heap, so return 1 for the a < b */
6389 486888 : if (sa->progress < sb->progress)
6390 436522 : return 1;
6391 50366 : else if (sa->progress == sb->progress)
6392 2298 : return 0;
6393 : else
6394 48068 : return -1;
6395 : }
6396 :
6397 : /*
6398 : * Initialize a writeback context, discarding potential previous state.
6399 : *
6400 : * *max_pending is a pointer instead of an immediate value, so the coalesce
6401 : * limits can easily changed by the GUC mechanism, and so calling code does
6402 : * not have to check the current configuration. A value of 0 means that no
6403 : * writeback control will be performed.
6404 : */
6405 : void
6406 5420 : WritebackContextInit(WritebackContext *context, int *max_pending)
6407 : {
6408 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6409 :
6410 5420 : context->max_pending = max_pending;
6411 5420 : context->nr_pending = 0;
6412 5420 : }
6413 :
6414 : /*
6415 : * Add buffer to list of pending writeback requests.
6416 : */
6417 : void
6418 1112324 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
6419 : BufferTag *tag)
6420 : {
6421 : PendingWriteback *pending;
6422 :
6423 : /*
6424 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
6425 : * point in tracking in that case.
6426 : */
6427 1112324 : if (io_direct_flags & IO_DIRECT_DATA ||
6428 1111280 : !enableFsync)
6429 1112324 : return;
6430 :
6431 : /*
6432 : * Add buffer to the pending writeback array, unless writeback control is
6433 : * disabled.
6434 : */
6435 0 : if (*wb_context->max_pending > 0)
6436 : {
6437 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6438 :
6439 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
6440 :
6441 0 : pending->tag = *tag;
6442 : }
6443 :
6444 : /*
6445 : * Perform pending flushes if the writeback limit is exceeded. This
6446 : * includes the case where previously an item has been added, but control
6447 : * is now disabled.
6448 : */
6449 0 : if (wb_context->nr_pending >= *wb_context->max_pending)
6450 0 : IssuePendingWritebacks(wb_context, io_context);
6451 : }
6452 :
6453 : #define ST_SORT sort_pending_writebacks
6454 : #define ST_ELEMENT_TYPE PendingWriteback
6455 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
6456 : #define ST_SCOPE static
6457 : #define ST_DEFINE
6458 : #include "lib/sort_template.h"
6459 :
6460 : /*
6461 : * Issue all pending writeback requests, previously scheduled with
6462 : * ScheduleBufferTagForWriteback, to the OS.
6463 : *
6464 : * Because this is only used to improve the OSs IO scheduling we try to never
6465 : * error out - it's just a hint.
6466 : */
6467 : void
6468 2138 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
6469 : {
6470 : instr_time io_start;
6471 : int i;
6472 :
6473 2138 : if (wb_context->nr_pending == 0)
6474 2138 : return;
6475 :
6476 : /*
6477 : * Executing the writes in-order can make them a lot faster, and allows to
6478 : * merge writeback requests to consecutive blocks into larger writebacks.
6479 : */
6480 0 : sort_pending_writebacks(wb_context->pending_writebacks,
6481 0 : wb_context->nr_pending);
6482 :
6483 0 : io_start = pgstat_prepare_io_time(track_io_timing);
6484 :
6485 : /*
6486 : * Coalesce neighbouring writes, but nothing else. For that we iterate
6487 : * through the, now sorted, array of pending flushes, and look forward to
6488 : * find all neighbouring (or identical) writes.
6489 : */
6490 0 : for (i = 0; i < wb_context->nr_pending; i++)
6491 : {
6492 : PendingWriteback *cur;
6493 : PendingWriteback *next;
6494 : SMgrRelation reln;
6495 : int ahead;
6496 : BufferTag tag;
6497 : RelFileLocator currlocator;
6498 0 : Size nblocks = 1;
6499 :
6500 0 : cur = &wb_context->pending_writebacks[i];
6501 0 : tag = cur->tag;
6502 0 : currlocator = BufTagGetRelFileLocator(&tag);
6503 :
6504 : /*
6505 : * Peek ahead, into following writeback requests, to see if they can
6506 : * be combined with the current one.
6507 : */
6508 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
6509 : {
6510 :
6511 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
6512 :
6513 : /* different file, stop */
6514 0 : if (!RelFileLocatorEquals(currlocator,
6515 0 : BufTagGetRelFileLocator(&next->tag)) ||
6516 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
6517 : break;
6518 :
6519 : /* ok, block queued twice, skip */
6520 0 : if (cur->tag.blockNum == next->tag.blockNum)
6521 0 : continue;
6522 :
6523 : /* only merge consecutive writes */
6524 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
6525 0 : break;
6526 :
6527 0 : nblocks++;
6528 0 : cur = next;
6529 : }
6530 :
6531 0 : i += ahead;
6532 :
6533 : /* and finally tell the kernel to write the data to storage */
6534 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
6535 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
6536 : }
6537 :
6538 : /*
6539 : * Assume that writeback requests are only issued for buffers containing
6540 : * blocks of permanent relations.
6541 : */
6542 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
6543 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
6544 :
6545 0 : wb_context->nr_pending = 0;
6546 : }
6547 :
6548 : /* ResourceOwner callbacks */
6549 :
6550 : static void
6551 30 : ResOwnerReleaseBufferIO(Datum res)
6552 : {
6553 30 : Buffer buffer = DatumGetInt32(res);
6554 :
6555 30 : AbortBufferIO(buffer);
6556 30 : }
6557 :
6558 : static char *
6559 0 : ResOwnerPrintBufferIO(Datum res)
6560 : {
6561 0 : Buffer buffer = DatumGetInt32(res);
6562 :
6563 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
6564 : }
6565 :
6566 : static void
6567 15162 : ResOwnerReleaseBufferPin(Datum res)
6568 : {
6569 15162 : Buffer buffer = DatumGetInt32(res);
6570 :
6571 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
6572 15162 : if (!BufferIsValid(buffer))
6573 0 : elog(ERROR, "bad buffer ID: %d", buffer);
6574 :
6575 15162 : if (BufferIsLocal(buffer))
6576 6066 : UnpinLocalBufferNoOwner(buffer);
6577 : else
6578 9096 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
6579 15162 : }
6580 :
6581 : static char *
6582 0 : ResOwnerPrintBufferPin(Datum res)
6583 : {
6584 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
6585 : }
6586 :
6587 : /*
6588 : * Helper function to evict unpinned buffer whose buffer header lock is
6589 : * already acquired.
6590 : */
6591 : static bool
6592 4278 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
6593 : {
6594 : uint32 buf_state;
6595 : bool result;
6596 :
6597 4278 : *buffer_flushed = false;
6598 :
6599 4278 : buf_state = pg_atomic_read_u32(&(desc->state));
6600 : Assert(buf_state & BM_LOCKED);
6601 :
6602 4278 : if ((buf_state & BM_VALID) == 0)
6603 : {
6604 0 : UnlockBufHdr(desc);
6605 0 : return false;
6606 : }
6607 :
6608 : /* Check that it's not pinned already. */
6609 4278 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
6610 : {
6611 0 : UnlockBufHdr(desc);
6612 0 : return false;
6613 : }
6614 :
6615 4278 : PinBuffer_Locked(desc); /* releases spinlock */
6616 :
6617 : /* If it was dirty, try to clean it once. */
6618 4278 : if (buf_state & BM_DIRTY)
6619 : {
6620 1940 : FlushUnlockedBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
6621 1940 : *buffer_flushed = true;
6622 : }
6623 :
6624 : /* This will return false if it becomes dirty or someone else pins it. */
6625 4278 : result = InvalidateVictimBuffer(desc);
6626 :
6627 4278 : UnpinBuffer(desc);
6628 :
6629 4278 : return result;
6630 : }
6631 :
6632 : /*
6633 : * Try to evict the current block in a shared buffer.
6634 : *
6635 : * This function is intended for testing/development use only!
6636 : *
6637 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
6638 : * particular block in mind, it might already have been replaced by some other
6639 : * block by the time this function runs. It's also unpinned on return, so the
6640 : * buffer might be occupied again by the time control is returned, potentially
6641 : * even by the same block. This inherent raciness without other interlocking
6642 : * makes the function unsuitable for non-testing usage.
6643 : *
6644 : * *buffer_flushed is set to true if the buffer was dirty and has been
6645 : * flushed, false otherwise. However, *buffer_flushed=true does not
6646 : * necessarily mean that we flushed the buffer, it could have been flushed by
6647 : * someone else.
6648 : *
6649 : * Returns true if the buffer was valid and it has now been made invalid.
6650 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
6651 : * or if the buffer becomes dirty again while we're trying to write it out.
6652 : */
6653 : bool
6654 280 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
6655 : {
6656 : BufferDesc *desc;
6657 :
6658 : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
6659 :
6660 : /* Make sure we can pin the buffer. */
6661 280 : ResourceOwnerEnlarge(CurrentResourceOwner);
6662 280 : ReservePrivateRefCountEntry();
6663 :
6664 280 : desc = GetBufferDescriptor(buf - 1);
6665 280 : LockBufHdr(desc);
6666 :
6667 280 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
6668 : }
6669 :
6670 : /*
6671 : * Try to evict all the shared buffers.
6672 : *
6673 : * This function is intended for testing/development use only! See
6674 : * EvictUnpinnedBuffer().
6675 : *
6676 : * The buffers_* parameters are mandatory and indicate the total count of
6677 : * buffers that:
6678 : * - buffers_evicted - were evicted
6679 : * - buffers_flushed - were flushed
6680 : * - buffers_skipped - could not be evicted
6681 : */
6682 : void
6683 2 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
6684 : int32 *buffers_skipped)
6685 : {
6686 2 : *buffers_evicted = 0;
6687 2 : *buffers_skipped = 0;
6688 2 : *buffers_flushed = 0;
6689 :
6690 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6691 : {
6692 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6693 : uint32 buf_state;
6694 : bool buffer_flushed;
6695 :
6696 32768 : CHECK_FOR_INTERRUPTS();
6697 :
6698 32768 : buf_state = pg_atomic_read_u32(&desc->state);
6699 32768 : if (!(buf_state & BM_VALID))
6700 28770 : continue;
6701 :
6702 3998 : ResourceOwnerEnlarge(CurrentResourceOwner);
6703 3998 : ReservePrivateRefCountEntry();
6704 :
6705 3998 : LockBufHdr(desc);
6706 :
6707 3998 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6708 3998 : (*buffers_evicted)++;
6709 : else
6710 0 : (*buffers_skipped)++;
6711 :
6712 3998 : if (buffer_flushed)
6713 1902 : (*buffers_flushed)++;
6714 : }
6715 2 : }
6716 :
6717 : /*
6718 : * Try to evict all the shared buffers containing provided relation's pages.
6719 : *
6720 : * This function is intended for testing/development use only! See
6721 : * EvictUnpinnedBuffer().
6722 : *
6723 : * The caller must hold at least AccessShareLock on the relation to prevent
6724 : * the relation from being dropped.
6725 : *
6726 : * The buffers_* parameters are mandatory and indicate the total count of
6727 : * buffers that:
6728 : * - buffers_evicted - were evicted
6729 : * - buffers_flushed - were flushed
6730 : * - buffers_skipped - could not be evicted
6731 : */
6732 : void
6733 2 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
6734 : int32 *buffers_flushed, int32 *buffers_skipped)
6735 : {
6736 : Assert(!RelationUsesLocalBuffers(rel));
6737 :
6738 2 : *buffers_skipped = 0;
6739 2 : *buffers_evicted = 0;
6740 2 : *buffers_flushed = 0;
6741 :
6742 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6743 : {
6744 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6745 32768 : uint32 buf_state = pg_atomic_read_u32(&(desc->state));
6746 : bool buffer_flushed;
6747 :
6748 32768 : CHECK_FOR_INTERRUPTS();
6749 :
6750 : /* An unlocked precheck should be safe and saves some cycles. */
6751 32768 : if ((buf_state & BM_VALID) == 0 ||
6752 54 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6753 32768 : continue;
6754 :
6755 : /* Make sure we can pin the buffer. */
6756 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
6757 0 : ReservePrivateRefCountEntry();
6758 :
6759 0 : buf_state = LockBufHdr(desc);
6760 :
6761 : /* recheck, could have changed without the lock */
6762 0 : if ((buf_state & BM_VALID) == 0 ||
6763 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6764 : {
6765 0 : UnlockBufHdr(desc);
6766 0 : continue;
6767 : }
6768 :
6769 0 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6770 0 : (*buffers_evicted)++;
6771 : else
6772 0 : (*buffers_skipped)++;
6773 :
6774 0 : if (buffer_flushed)
6775 0 : (*buffers_flushed)++;
6776 : }
6777 2 : }
6778 :
6779 : /*
6780 : * Generic implementation of the AIO handle staging callback for readv/writev
6781 : * on local/shared buffers.
6782 : *
6783 : * Each readv/writev can target multiple buffers. The buffers have already
6784 : * been registered with the IO handle.
6785 : *
6786 : * To make the IO ready for execution ("staging"), we need to ensure that the
6787 : * targeted buffers are in an appropriate state while the IO is ongoing. For
6788 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
6789 : * in this backend could lead to this backend's buffer pin being released as
6790 : * part of error handling, which in turn could lead to the buffer being
6791 : * replaced while IO is ongoing.
6792 : */
6793 : static pg_attribute_always_inline void
6794 2513092 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
6795 : {
6796 : uint64 *io_data;
6797 : uint8 handle_data_len;
6798 : PgAioWaitRef io_ref;
6799 2513092 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
6800 :
6801 2513092 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
6802 :
6803 2513092 : pgaio_io_get_wref(ioh, &io_ref);
6804 :
6805 : /* iterate over all buffers affected by the vectored readv/writev */
6806 5360866 : for (int i = 0; i < handle_data_len; i++)
6807 : {
6808 2847774 : Buffer buffer = (Buffer) io_data[i];
6809 2847774 : BufferDesc *buf_hdr = is_temp ?
6810 16818 : GetLocalBufferDescriptor(-buffer - 1)
6811 2847774 : : GetBufferDescriptor(buffer - 1);
6812 : uint32 buf_state;
6813 :
6814 : /*
6815 : * Check that all the buffers are actually ones that could conceivably
6816 : * be done in one IO, i.e. are sequential. This is the last
6817 : * buffer-aware code before IO is actually executed and confusion
6818 : * about which buffers are targeted by IO can be hard to debug, making
6819 : * it worth doing extra-paranoid checks.
6820 : */
6821 2847774 : if (i == 0)
6822 2513092 : first = buf_hdr->tag;
6823 : else
6824 : {
6825 : Assert(buf_hdr->tag.relNumber == first.relNumber);
6826 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
6827 : }
6828 :
6829 2847774 : if (is_temp)
6830 16818 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
6831 : else
6832 2830956 : buf_state = LockBufHdr(buf_hdr);
6833 :
6834 : /* verify the buffer is in the expected state */
6835 : Assert(buf_state & BM_TAG_VALID);
6836 : if (is_write)
6837 : {
6838 : Assert(buf_state & BM_VALID);
6839 : Assert(buf_state & BM_DIRTY);
6840 : }
6841 : else
6842 : {
6843 : Assert(!(buf_state & BM_VALID));
6844 : Assert(!(buf_state & BM_DIRTY));
6845 : }
6846 :
6847 : /* temp buffers don't use BM_IO_IN_PROGRESS */
6848 2847774 : if (!is_temp)
6849 : Assert(buf_state & BM_IO_IN_PROGRESS);
6850 :
6851 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
6852 :
6853 : /*
6854 : * Reflect that the buffer is now owned by the AIO subsystem.
6855 : *
6856 : * For local buffers: This can't be done just via LocalRefCount, as
6857 : * one might initially think, as this backend could error out while
6858 : * AIO is still in progress, releasing all the pins by the backend
6859 : * itself.
6860 : *
6861 : * This pin is released again in TerminateBufferIO().
6862 : */
6863 2847774 : buf_hdr->io_wref = io_ref;
6864 :
6865 2847774 : if (is_temp)
6866 : {
6867 16818 : buf_state += BUF_REFCOUNT_ONE;
6868 16818 : pg_atomic_unlocked_write_u32(&buf_hdr->state, buf_state);
6869 : }
6870 : else
6871 2830956 : UnlockBufHdrExt(buf_hdr, buf_state, 0, 0, 1);
6872 :
6873 : /*
6874 : * Ensure the content lock that prevents buffer modifications while
6875 : * the buffer is being written out is not released early due to an
6876 : * error.
6877 : */
6878 2847774 : if (is_write && !is_temp)
6879 : {
6880 : LWLock *content_lock;
6881 :
6882 0 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
6883 :
6884 : Assert(LWLockHeldByMe(content_lock));
6885 :
6886 : /*
6887 : * Lock is now owned by AIO subsystem.
6888 : */
6889 0 : LWLockDisown(content_lock);
6890 : }
6891 :
6892 : /*
6893 : * Stop tracking this buffer via the resowner - the AIO system now
6894 : * keeps track.
6895 : */
6896 2847774 : if (!is_temp)
6897 2830956 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
6898 : }
6899 2513092 : }
6900 :
6901 : /*
6902 : * Decode readv errors as encoded by buffer_readv_encode_error().
6903 : */
6904 : static inline void
6905 698 : buffer_readv_decode_error(PgAioResult result,
6906 : bool *zeroed_any,
6907 : bool *ignored_any,
6908 : uint8 *zeroed_or_error_count,
6909 : uint8 *checkfail_count,
6910 : uint8 *first_off)
6911 : {
6912 698 : uint32 rem_error = result.error_data;
6913 :
6914 : /* see static asserts in buffer_readv_encode_error */
6915 : #define READV_COUNT_BITS 7
6916 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
6917 :
6918 698 : *zeroed_any = rem_error & 1;
6919 698 : rem_error >>= 1;
6920 :
6921 698 : *ignored_any = rem_error & 1;
6922 698 : rem_error >>= 1;
6923 :
6924 698 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
6925 698 : rem_error >>= READV_COUNT_BITS;
6926 :
6927 698 : *checkfail_count = rem_error & READV_COUNT_MASK;
6928 698 : rem_error >>= READV_COUNT_BITS;
6929 :
6930 698 : *first_off = rem_error & READV_COUNT_MASK;
6931 698 : rem_error >>= READV_COUNT_BITS;
6932 698 : }
6933 :
6934 : /*
6935 : * Helper to encode errors for buffer_readv_complete()
6936 : *
6937 : * Errors are encoded as follows:
6938 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
6939 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
6940 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
6941 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
6942 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
6943 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
6944 : * checksum
6945 : */
6946 : static inline void
6947 384 : buffer_readv_encode_error(PgAioResult *result,
6948 : bool is_temp,
6949 : bool zeroed_any,
6950 : bool ignored_any,
6951 : uint8 error_count,
6952 : uint8 zeroed_count,
6953 : uint8 checkfail_count,
6954 : uint8 first_error_off,
6955 : uint8 first_zeroed_off,
6956 : uint8 first_ignored_off)
6957 : {
6958 :
6959 384 : uint8 shift = 0;
6960 384 : uint8 zeroed_or_error_count =
6961 : error_count > 0 ? error_count : zeroed_count;
6962 : uint8 first_off;
6963 :
6964 : StaticAssertStmt(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
6965 : "PG_IOV_MAX is bigger than reserved space for error data");
6966 : StaticAssertStmt((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
6967 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
6968 :
6969 : /*
6970 : * We only have space to encode one offset - but luckily that's good
6971 : * enough. If there is an error, the error is the interesting offset, same
6972 : * with a zeroed buffer vs an ignored buffer.
6973 : */
6974 384 : if (error_count > 0)
6975 188 : first_off = first_error_off;
6976 196 : else if (zeroed_count > 0)
6977 160 : first_off = first_zeroed_off;
6978 : else
6979 36 : first_off = first_ignored_off;
6980 :
6981 : Assert(!zeroed_any || error_count == 0);
6982 :
6983 384 : result->error_data = 0;
6984 :
6985 384 : result->error_data |= zeroed_any << shift;
6986 384 : shift += 1;
6987 :
6988 384 : result->error_data |= ignored_any << shift;
6989 384 : shift += 1;
6990 :
6991 384 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
6992 384 : shift += READV_COUNT_BITS;
6993 :
6994 384 : result->error_data |= ((uint32) checkfail_count) << shift;
6995 384 : shift += READV_COUNT_BITS;
6996 :
6997 384 : result->error_data |= ((uint32) first_off) << shift;
6998 384 : shift += READV_COUNT_BITS;
6999 :
7000 384 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
7001 : PGAIO_HCB_SHARED_BUFFER_READV;
7002 :
7003 384 : if (error_count > 0)
7004 188 : result->status = PGAIO_RS_ERROR;
7005 : else
7006 196 : result->status = PGAIO_RS_WARNING;
7007 :
7008 : /*
7009 : * The encoding is complicated enough to warrant cross-checking it against
7010 : * the decode function.
7011 : */
7012 : #ifdef USE_ASSERT_CHECKING
7013 : {
7014 : bool zeroed_any_2,
7015 : ignored_any_2;
7016 : uint8 zeroed_or_error_count_2,
7017 : checkfail_count_2,
7018 : first_off_2;
7019 :
7020 : buffer_readv_decode_error(*result,
7021 : &zeroed_any_2, &ignored_any_2,
7022 : &zeroed_or_error_count_2,
7023 : &checkfail_count_2,
7024 : &first_off_2);
7025 : Assert(zeroed_any == zeroed_any_2);
7026 : Assert(ignored_any == ignored_any_2);
7027 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
7028 : Assert(checkfail_count == checkfail_count_2);
7029 : Assert(first_off == first_off_2);
7030 : }
7031 : #endif
7032 :
7033 : #undef READV_COUNT_BITS
7034 : #undef READV_COUNT_MASK
7035 384 : }
7036 :
7037 : /*
7038 : * Helper for AIO readv completion callbacks, supporting both shared and temp
7039 : * buffers. Gets called once for each buffer in a multi-page read.
7040 : */
7041 : static pg_attribute_always_inline void
7042 2572312 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
7043 : uint8 flags, bool failed, bool is_temp,
7044 : bool *buffer_invalid,
7045 : bool *failed_checksum,
7046 : bool *ignored_checksum,
7047 : bool *zeroed_buffer)
7048 : {
7049 2572312 : BufferDesc *buf_hdr = is_temp ?
7050 16818 : GetLocalBufferDescriptor(-buffer - 1)
7051 2572312 : : GetBufferDescriptor(buffer - 1);
7052 2572312 : BufferTag tag = buf_hdr->tag;
7053 2572312 : char *bufdata = BufferGetBlock(buffer);
7054 : uint32 set_flag_bits;
7055 : int piv_flags;
7056 :
7057 : /* check that the buffer is in the expected state for a read */
7058 : #ifdef USE_ASSERT_CHECKING
7059 : {
7060 : uint32 buf_state = pg_atomic_read_u32(&buf_hdr->state);
7061 :
7062 : Assert(buf_state & BM_TAG_VALID);
7063 : Assert(!(buf_state & BM_VALID));
7064 : /* temp buffers don't use BM_IO_IN_PROGRESS */
7065 : if (!is_temp)
7066 : Assert(buf_state & BM_IO_IN_PROGRESS);
7067 : Assert(!(buf_state & BM_DIRTY));
7068 : }
7069 : #endif
7070 :
7071 2572312 : *buffer_invalid = false;
7072 2572312 : *failed_checksum = false;
7073 2572312 : *ignored_checksum = false;
7074 2572312 : *zeroed_buffer = false;
7075 :
7076 : /*
7077 : * We ask PageIsVerified() to only log the message about checksum errors,
7078 : * as the completion might be run in any backend (or IO workers). We will
7079 : * report checksum errors in buffer_readv_report().
7080 : */
7081 2572312 : piv_flags = PIV_LOG_LOG;
7082 :
7083 : /* the local zero_damaged_pages may differ from the definer's */
7084 2572312 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
7085 76 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
7086 :
7087 : /* Check for garbage data. */
7088 2572312 : if (!failed)
7089 : {
7090 : /*
7091 : * If the buffer is not currently pinned by this backend, e.g. because
7092 : * we're completing this IO after an error, the buffer data will have
7093 : * been marked as inaccessible when the buffer was unpinned. The AIO
7094 : * subsystem holds a pin, but that doesn't prevent the buffer from
7095 : * having been marked as inaccessible. The completion might also be
7096 : * executed in a different process.
7097 : */
7098 : #ifdef USE_VALGRIND
7099 : if (!BufferIsPinned(buffer))
7100 : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
7101 : #endif
7102 :
7103 2572254 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
7104 : failed_checksum))
7105 : {
7106 192 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
7107 : {
7108 92 : memset(bufdata, 0, BLCKSZ);
7109 92 : *zeroed_buffer = true;
7110 : }
7111 : else
7112 : {
7113 100 : *buffer_invalid = true;
7114 : /* mark buffer as having failed */
7115 100 : failed = true;
7116 : }
7117 : }
7118 2572062 : else if (*failed_checksum)
7119 24 : *ignored_checksum = true;
7120 :
7121 : /* undo what we did above */
7122 : #ifdef USE_VALGRIND
7123 : if (!BufferIsPinned(buffer))
7124 : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
7125 : #endif
7126 :
7127 : /*
7128 : * Immediately log a message about the invalid page, but only to the
7129 : * server log. The reason to do so immediately is that this may be
7130 : * executed in a different backend than the one that originated the
7131 : * request. The reason to do so immediately is that the originator
7132 : * might not process the query result immediately (because it is busy
7133 : * doing another part of query processing) or at all (e.g. if it was
7134 : * cancelled or errored out due to another IO also failing). The
7135 : * definer of the IO will emit an ERROR or WARNING when processing the
7136 : * IO's results
7137 : *
7138 : * To avoid duplicating the code to emit these log messages, we reuse
7139 : * buffer_readv_report().
7140 : */
7141 2572254 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
7142 : {
7143 216 : PgAioResult result_one = {0};
7144 :
7145 216 : buffer_readv_encode_error(&result_one, is_temp,
7146 216 : *zeroed_buffer,
7147 216 : *ignored_checksum,
7148 216 : *buffer_invalid,
7149 216 : *zeroed_buffer ? 1 : 0,
7150 216 : *failed_checksum ? 1 : 0,
7151 : buf_off, buf_off, buf_off);
7152 216 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
7153 : }
7154 : }
7155 :
7156 : /* Terminate I/O and set BM_VALID. */
7157 2572312 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
7158 2572312 : if (is_temp)
7159 16818 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
7160 : else
7161 2555494 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
7162 :
7163 : /*
7164 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
7165 : * callback may not be executed in the same backend that called
7166 : * BUFFER_READ_START. The alternative would be to defer calling the
7167 : * tracepoint to a later point (e.g. the local completion callback for
7168 : * shared buffer reads), which seems even less helpful.
7169 : */
7170 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
7171 : tag.blockNum,
7172 : tag.spcOid,
7173 : tag.dbOid,
7174 : tag.relNumber,
7175 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
7176 : false);
7177 2572312 : }
7178 :
7179 : /*
7180 : * Perform completion handling of a single AIO read. This read may cover
7181 : * multiple blocks / buffers.
7182 : *
7183 : * Shared between shared and local buffers, to reduce code duplication.
7184 : */
7185 : static pg_attribute_always_inline PgAioResult
7186 2303886 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7187 : uint8 cb_data, bool is_temp)
7188 : {
7189 2303886 : PgAioResult result = prior_result;
7190 2303886 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7191 2303886 : uint8 first_error_off = 0;
7192 2303886 : uint8 first_zeroed_off = 0;
7193 2303886 : uint8 first_ignored_off = 0;
7194 2303886 : uint8 error_count = 0;
7195 2303886 : uint8 zeroed_count = 0;
7196 2303886 : uint8 ignored_count = 0;
7197 2303886 : uint8 checkfail_count = 0;
7198 : uint64 *io_data;
7199 : uint8 handle_data_len;
7200 :
7201 : if (is_temp)
7202 : {
7203 : Assert(td->smgr.is_temp);
7204 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
7205 : }
7206 : else
7207 : Assert(!td->smgr.is_temp);
7208 :
7209 : /*
7210 : * Iterate over all the buffers affected by this IO and call the
7211 : * per-buffer completion function for each buffer.
7212 : */
7213 2303886 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
7214 4876198 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
7215 : {
7216 2572312 : Buffer buf = io_data[buf_off];
7217 : bool failed;
7218 2572312 : bool failed_verification = false;
7219 2572312 : bool failed_checksum = false;
7220 2572312 : bool zeroed_buffer = false;
7221 2572312 : bool ignored_checksum = false;
7222 :
7223 : Assert(BufferIsValid(buf));
7224 :
7225 : /*
7226 : * If the entire I/O failed on a lower-level, each buffer needs to be
7227 : * marked as failed. In case of a partial read, the first few buffers
7228 : * may be ok.
7229 : */
7230 2572312 : failed =
7231 2572312 : prior_result.status == PGAIO_RS_ERROR
7232 2572312 : || prior_result.result <= buf_off;
7233 :
7234 2572312 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
7235 : &failed_verification,
7236 : &failed_checksum,
7237 : &ignored_checksum,
7238 : &zeroed_buffer);
7239 :
7240 : /*
7241 : * Track information about the number of different kinds of error
7242 : * conditions across all pages, as there can be multiple pages failing
7243 : * verification as part of one IO.
7244 : */
7245 2572312 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
7246 88 : first_error_off = buf_off;
7247 2572312 : if (zeroed_buffer && zeroed_count++ == 0)
7248 68 : first_zeroed_off = buf_off;
7249 2572312 : if (ignored_checksum && ignored_count++ == 0)
7250 20 : first_ignored_off = buf_off;
7251 2572312 : if (failed_checksum)
7252 64 : checkfail_count++;
7253 : }
7254 :
7255 : /*
7256 : * If the smgr read succeeded [partially] and page verification failed for
7257 : * some of the pages, adjust the IO's result state appropriately.
7258 : */
7259 2303886 : if (prior_result.status != PGAIO_RS_ERROR &&
7260 2303780 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
7261 : {
7262 168 : buffer_readv_encode_error(&result, is_temp,
7263 : zeroed_count > 0, ignored_count > 0,
7264 : error_count, zeroed_count, checkfail_count,
7265 : first_error_off, first_zeroed_off,
7266 : first_ignored_off);
7267 168 : pgaio_result_report(result, td, DEBUG1);
7268 : }
7269 :
7270 : /*
7271 : * For shared relations this reporting is done in
7272 : * shared_buffer_readv_complete_local().
7273 : */
7274 2303886 : if (is_temp && checkfail_count > 0)
7275 4 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7276 : checkfail_count);
7277 :
7278 2303886 : return result;
7279 : }
7280 :
7281 : /*
7282 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
7283 : * aio_local_buffer_readv_cb.
7284 : *
7285 : * The error is encoded / decoded in buffer_readv_encode_error() /
7286 : * buffer_readv_decode_error().
7287 : */
7288 : static void
7289 544 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
7290 : int elevel)
7291 : {
7292 544 : int nblocks = td->smgr.nblocks;
7293 544 : BlockNumber first = td->smgr.blockNum;
7294 544 : BlockNumber last = first + nblocks - 1;
7295 544 : ProcNumber errProc =
7296 544 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
7297 : RelPathStr rpath =
7298 544 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
7299 : bool zeroed_any,
7300 : ignored_any;
7301 : uint8 zeroed_or_error_count,
7302 : checkfail_count,
7303 : first_off;
7304 : uint8 affected_count;
7305 : const char *msg_one,
7306 : *msg_mult,
7307 : *det_mult,
7308 : *hint_mult;
7309 :
7310 544 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
7311 : &zeroed_or_error_count,
7312 : &checkfail_count,
7313 : &first_off);
7314 :
7315 : /*
7316 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
7317 : * special case, it's too irregular to be emitted the same way as the
7318 : * other cases.
7319 : */
7320 544 : if (zeroed_any && ignored_any)
7321 : {
7322 : Assert(zeroed_any && ignored_any);
7323 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
7324 : Assert(result.status != PGAIO_RS_ERROR);
7325 8 : affected_count = zeroed_or_error_count;
7326 :
7327 8 : ereport(elevel,
7328 : errcode(ERRCODE_DATA_CORRUPTED),
7329 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
7330 : affected_count, checkfail_count, first, last, rpath.str),
7331 : affected_count > 1 ?
7332 : errdetail("Block %u held the first zeroed page.",
7333 : first + first_off) : 0,
7334 : errhint_plural("See server log for details about the other %d invalid block.",
7335 : "See server log for details about the other %d invalid blocks.",
7336 : affected_count + checkfail_count - 1,
7337 : affected_count + checkfail_count - 1));
7338 8 : return;
7339 : }
7340 :
7341 : /*
7342 : * The other messages are highly repetitive. To avoid duplicating a long
7343 : * and complicated ereport(), gather the translated format strings
7344 : * separately and then do one common ereport.
7345 : */
7346 536 : if (result.status == PGAIO_RS_ERROR)
7347 : {
7348 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
7349 272 : affected_count = zeroed_or_error_count;
7350 272 : msg_one = _("invalid page in block %u of relation \"%s\"");
7351 272 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
7352 272 : det_mult = _("Block %u held the first invalid page.");
7353 272 : hint_mult = _("See server log for the other %u invalid block(s).");
7354 : }
7355 264 : else if (zeroed_any && !ignored_any)
7356 : {
7357 216 : affected_count = zeroed_or_error_count;
7358 216 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
7359 216 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
7360 216 : det_mult = _("Block %u held the first zeroed page.");
7361 216 : hint_mult = _("See server log for the other %u zeroed block(s).");
7362 : }
7363 48 : else if (!zeroed_any && ignored_any)
7364 : {
7365 48 : affected_count = checkfail_count;
7366 48 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
7367 48 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
7368 48 : det_mult = _("Block %u held the first ignored page.");
7369 48 : hint_mult = _("See server log for the other %u ignored block(s).");
7370 : }
7371 : else
7372 0 : pg_unreachable();
7373 :
7374 536 : ereport(elevel,
7375 : errcode(ERRCODE_DATA_CORRUPTED),
7376 : affected_count == 1 ?
7377 : errmsg_internal(msg_one, first + first_off, rpath.str) :
7378 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
7379 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
7380 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
7381 : }
7382 :
7383 : static void
7384 2509480 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7385 : {
7386 2509480 : buffer_stage_common(ioh, false, false);
7387 2509480 : }
7388 :
7389 : static PgAioResult
7390 2300274 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7391 : uint8 cb_data)
7392 : {
7393 2300274 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
7394 : }
7395 :
7396 : /*
7397 : * We need a backend-local completion callback for shared buffers, to be able
7398 : * to report checksum errors correctly. Unfortunately that can only safely
7399 : * happen if the reporting backend has previously called
7400 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
7401 : * the backend that started the IO. Hence this callback.
7402 : */
7403 : static PgAioResult
7404 2509480 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
7405 : uint8 cb_data)
7406 : {
7407 : bool zeroed_any,
7408 : ignored_any;
7409 : uint8 zeroed_or_error_count,
7410 : checkfail_count,
7411 : first_off;
7412 :
7413 2509480 : if (prior_result.status == PGAIO_RS_OK)
7414 2509326 : return prior_result;
7415 :
7416 154 : buffer_readv_decode_error(prior_result,
7417 : &zeroed_any,
7418 : &ignored_any,
7419 : &zeroed_or_error_count,
7420 : &checkfail_count,
7421 : &first_off);
7422 :
7423 154 : if (checkfail_count)
7424 : {
7425 48 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7426 :
7427 48 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7428 : checkfail_count);
7429 : }
7430 :
7431 154 : return prior_result;
7432 : }
7433 :
7434 : static void
7435 3612 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7436 : {
7437 3612 : buffer_stage_common(ioh, false, true);
7438 3612 : }
7439 :
7440 : static PgAioResult
7441 3612 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7442 : uint8 cb_data)
7443 : {
7444 3612 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
7445 : }
7446 :
7447 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7448 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
7449 : .stage = shared_buffer_readv_stage,
7450 : .complete_shared = shared_buffer_readv_complete,
7451 : /* need a local callback to report checksum failures */
7452 : .complete_local = shared_buffer_readv_complete_local,
7453 : .report = buffer_readv_report,
7454 : };
7455 :
7456 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7457 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
7458 : .stage = local_buffer_readv_stage,
7459 :
7460 : /*
7461 : * Note that this, in contrast to the shared_buffers case, uses
7462 : * complete_local, as only the issuing backend has access to the required
7463 : * datastructures. This is important in case the IO completion may be
7464 : * consumed incidentally by another backend.
7465 : */
7466 : .complete_local = local_buffer_readv_complete,
7467 : .report = buffer_readv_report,
7468 : };
|
